Mining Imprecise Discriminative Molecular Fragments

Michael R. Berthold

We present an algorithm to find imprecise fragments in a set of molecules that help to discriminate between different classes of, for instance, activity in a drug discovery context. Instead of carrying out a brute-force search, our method generates fragments by embedding them in all appropriate molecules in parallel and prunes the search tree based on a local order of the atoms and bonds, which results in substantially faster search by eliminating the need for frequent, computationally expensive reembeddings and by suppressing redundant search. An extension of the search algorithm allows finding fragments that also incorporate chemical expert knowledge about structural similarities that can be tolerated. We demonstrate the usefulness of our algorithm by demonstrating the discovery of activity-related groups of chemical compounds in the National Cancer Institute's HIV and Cancer-screening datasets.

Mining Fuzzy Association Rules: an Overview

Miguel Delgado

The theory of Fuzzy Sets has been recognized as a suitable tool to model several kinds of patterns that can hold in data. Particularly Fuzzy Sets Theory has been shown to be a very useful tool in Data Mining to representing a natural and human way the so called Association Rules The objective of this paper is to present a revision of the most relevant results about the use of Fuzzy Sets in Data Mining, specially in relation with the discovery of Association Rules. First of all we will introduce the basic concepts of Data Mining to justify the need of Fuzzy Sets Theory. A historical revision on developments in this field is made too Next we will present our researches about Fuzzy Association Rules, starting with the formulation of a model to discover association rules among items in a (crisp) set of fuzzy transactions. This general model can be particularized in several ways; each particular instance corresponds to a different kind of pattern and/or repository of data. We describe some applications of this scheme, paying special attention to the discovery of fuzzy association rules in relational databases Paper finishes with some suggestions about future researches and problems to be solved.

Principal Component and Self-aggregation Clustering of Gene Expressions and Protein Interactions
Chris Ding

Cluster analysis partitions data points into disjoint groups to quickly gain first order knowledge from data. Recently we developed several principal component based clustering algorithms with well defined clustering objective functions. For the widely used K-means clustering we proved that the continuous solution of the cluster membership indicator vector is the principal component[1], leading to effective implementations of K-means clustering. For spectral graph clustering, we proved that scaled principal components form a dynamic process of self-aggregation in which data objects move towards each other to form clusters, revealing the inherent pattern of similarity[2]. The MinMaxCut spectral method follows a min-max clustering principle that the between-cluster associations are minimized, while the within-cluster associations are maximized.

DNA microarrays can simultaneously monitor expression levels of thousands of genes. Expression profiles of patients tissue samples provide molecular rather than morphological signature of cancer. We apply data clustering to expression profiles of human Lymphoma and demonstrate that data clustering can effective discover different pathological stages of B-cell lymphoma thus providing an effective diagnostics methodology[3].

Proteins carry out many cellular functions such as metabolism, communication, growth. Systematic identification of these protein complexes provide essential knowledge linking proteome dynamics to cellular function and phenotype. We describe a unified representation of protein complex dataset based on bipartite graphs, from which protein-protein interactions and protein-complex - protein-complex assocations are deducted naturally. Apply data clustering to the protein-protein network we obtain statistically significant protein clusters. Apply data clustering to the protein-complex - protein-complex association network, we obtain protein supercomplexes that provide systems-level understanding of cellular processes through the concepts from Gene Ontology[4].

[1] C. Ding and X. He. Principal component analysis and K-means clustering. LBNL Tech Report 52983. June 6, 2003.
[2] C. Ding, X. He, H. Zha, and H. Simon. Unsupervised learning: self-aggregation in scaled principal component space. Proc. of Principles of Data Mining and Knowledge Discovery (PKDD 2002), pp. 112-124. Aug. 2002.
[3] C. Ding. Analysis of gene expression profiles: class discovery and leaf ordering. Proc. of Conf. on Research in Comp.Mol.Bio (RECOMB 2002), pp.127-136. April 2002.
[4] C. Ding, X. He, R. Meraz, S. Holbrook. A unified representation for multi-protein complex data for modeling protein interactions. LBNL Tech Report 52984. To appear in Proteins: Structure, Function, and Genetics.

Augmented Fuzzy Cognitive Maps based on Case Based Reasoning for Decisions in Medical Informatics

Voula C. Georgopoulos

This paper presents the Soft Computing technique of Fuzzy Cognitive Maps (FCMs), which is a knowledge-based methodology suitable to describe and model complex systems and handle information from an abstract point of view [1]. Soft computing techniques such as FCMs have been successfully used to model complex systems that involve discipline factors, states, variables, input, output, events and trends. These modeling techniques can integrate and include in the decision-making process the partial influence of controversial factors, can take under consideration causal effect among factors and evaluate the influence from different sources, factors and other characteristics using fuzzy logic reasoning. Thus, such methods are ideal for the development of Decision Support systems in Medical Informatics, because in this application area humans use decision approaches, such as in differential diagnosis, for example, mainly based on fuzzy factors.Some of these factors are complementary, others similar and others conflicting, which are all taken into consideration when a decision is reached [2, 3, 4]. The involved factors have different degrees of importance in determining (or influencing) the decision, and moreover the fact that they can be additive and/or conflicting increases the complexity of the problem and the vagueness of the decision.

Fuzzy Cognitive Maps develop a behavioral model of the system by exploiting the experience and knowledge of experts due to the way that they are developed. Fuzzy Cognitive Maps applicability in modeling complex systems has been successfully introduced [5]. FCMs originated from the synergism of Fuzzy Logic and Neural Networks, taking advantage of both theories. An FCM is a signed fuzzy graph with feedback, consisting of concepts-nodes and weighted interconnections. Nodes of the graph stand for concepts that are used to describe main behavioral characteristics of the modeled system. Nodes are connected by signed and fuzzy weighted arcs representing the cause and effect relationship existing among concepts. Thus, an FCM is a fuzzy-graph structure, which allows systematic causal propagation, in particular forward and backward chaining [6]. Fuzzy Cognitive Maps have been successfully used to develop a Decision Support System for differential diagnosis [7], to determine the success of radiation therapy process estimating the final dose delivered to the target volume [8] and many other application areas.

When a FCM is used as part of a decision-making system, some of the concepts are considered output nodes, which correspond to each of the possible decisions. In general, the FCM could converge to a fixed point, limit cycle, or chaotic attractor, but when FCMs are used for the decision making process, it is desirable to converge to a region corresponding to the selection of one decision. Therefore, the output nodes must "compete" against each other in order for only one of them to dominate and be considered the decision. Here a new idea is proposed for achieving this "competition" between concepts. The interaction of each of these output nodes with the other output nodes should be strongly inhibitory. This implies that the higher the value of a given node, this should lead to a lowering of the value of competing nodes. In such a case the FCM, referred to as a Competitive Fuzzy Cognitive Map (CFCM) will always converge to a fixed appropriate region.

In this research work we further develop CFCMs by combining CFCMs with methods and approaches that have been used for Case-Based Reasoning (CBR). This is a successful methodology for managing implicit knowledge, which has also been used in medical informatics. CBRs embed a considerable amount of previous solved instances of problems (called cases). The problem solving experience is explicitly taken into account by storing past cases in database (case base), and by suitably retrieving them when a new problem has to be tackled [9]. It simply solves new cases by similarity of the solution of the old cases stored in its case-base rather than using some derivative representation, as is done for example in adaptive-type methodologies. But, if the new case has no match with the stored cases, CBR has no solution. Similarly, to FCMs and CFCMs, CBRs have been applied in medical diagnosis and patient treatment outcomes. Despite the limitations of CBRs they are usually assumed to have a certain degree of richness of stored knowledge, and a certain degree of complexity due to the way they are organized.

Even though successful medical Decision Support CFCMs have been developed [7], [10] there are situations where the patient data to be input into the system presents a very rare configuration of symptoms where most of the nodes of the CFCM would not be active. In other words, for example, although the CFCM-Model of a Medical Decision Support System has been designed to include all possible symptoms and causative factors (nodes-concepts) and the relationship between them (weights) for some medical condition, in a particular situation very few symptoms are available and are taken into consideration. Thus, in such a diagnosis or prognosis model Decision Support FCM, the decision would be made only using a very small subset of the concepts of the entire system. Such a system could lead to either an erroneous decision or difficulty in reaching stability since the weighting of the active nodes reflects only a small amount of the experts' stored knowledge. Using a CBR-Augmented CFCM Decision support system, in such situations, the decisions support system would draw upon cases that are maximally similar according to distance measures and would use the CBR subsystem to generate a sub-CFCM emphasizing the nodes activated by the patient data and thus redistributing the causal weightings between the concept-nodes.

The advantage of CBR-augmented CFCMs lies in the ability to represent rare occurrences of medical conditions/symptoms, which may not be adequately represented in an FCM due to its design methodology, which is dependent on human experts and learning algorithms.

In this paper, we describe the algorithm that generates the CBR-augmented CFCMs and present their use and applicability for a Medical Decision System for the Speech and Language Impaired. Specifically the CBR-augmented CFCMs is proposed for Differential Diagnosis of Specific Language Impairment from Dyslexia and Autism, as an extension of an CFCM that has been developed [7], [10].

[1] B. Kosko, “Fuzzy Cognitive Maps, International J. of Man-Machine studies, vol.24, pp. 65-75 (1986).
[2] Kasabov N. Foundations of Neural Networks, Fuzzy Systems and Knowledge Engineering, Cambridge, Massachusetts: MIT Press (1996).
[3] Kasabov, N. Decision support systems and expert systems, in: M. Arbib (ed) Handbook of brain study and neural networks, Cambridge, Massachusetts: MIT Press (2002).
[4] J. Zeleznikow and J. Nolan, Using Soft computing to build real world intelligent decision support systems in uncertain domains, Decision Support Systems, Vol. 31, pp. 263-285 (2001).
[5] C. D. Stylios and P.P. Groumpos, V. C Georgopoulos "An Fuzzy Cognitive Maps Approach to Process Control Systems" J.Advanced Computational Intelligence, Vol. 3, No5, pp.1-9, (1999)
[6] C.D. Stylios and P.P. Groumpos, "Fuzzy Cognitive Maps in Modeling Supervisory Control Systems" Journal of Intelligent & Fuzzy Systems, Vol 8. No 2. pp. 83-98 (2000).
[7] V.C. Georgopoulos, G.A. Malandraki and C. D. Stylios "A Fuzzy Cognitive Map Approach To Differential Diagnosis of Specific Language Impairment" Journal of Artificial Intelligence in Medicine, Vol. 29, No.3, pp. 261-278, (2003).
[8] E. Papageorgiou, C. Stylios and P. Groumpos "An Integrated Two-level Hierarchical System for Decision Making in Radiation Therapy using Fuzzy Cognitive" IEEE Transactions on Biomedical Engineering vol. 50, No 12 (December 2003).
[9] J.B. Noh, K.C Lee, J.K. Kim, J.K. Lee and S.H. Kim, "A case-based reasoning approach to cognitive map-driven tacit knowledge management" Experts Systems with Applications, Vol 19, pp. 249-259 (2000).br> [10] V.C. Georgopoulos and C. D. Stylios, "Competitive Fuzzy Cognitive Maps for Decision-Making Systems", presented at Comcon 8, The 8th International Conference on Advances in Communication and Control: Telecommunications/Signal Processing, Crete, Greece, June 25-29 (2001).

Soft Computing Methods for Drug Dosing in Chronic Anemia

Adam Gaweda

Drug delivery in chronic therapy is widely recognized as a complex control problem. The non-stationary character of the human body, as a result of tolerance to the drug and other pharmacokinetic parameters changing over time, requires intermittent therapy adjustment. Furthermore, it is virtually impossible to regularly access all the relevant variables to accurately represent a system as complex as human body. Hence, the development of a parametric model becomes difficult and the need to deal with imprecision arises. This presentation demonstrates how these two aspects are addressed by employing Soft Computing methods in long-term dosing of erythropoietin, a drug which is used for the treatment of patients with chronic anemia. Artificial Neural Network-based adaptive controller for erythropoietin dose adjustment will be presented and evaluated with respect to the currently used protocol. A modified Fuzzy Rule-Based System, capable of handling imprecise input and output information, will be introduced and its application to predict patients' response to erythropoietin will be demonstrated.

Collaborative Agent Systems for Distributed Automation

Dilip B. Kotak, William A. Gruver

Computer and communication technologies are very rapidly shrinking the world. Manufacturing and supply chains, the service industry and the infrastructure are becoming highly distributed and also highly integrated. This creates new challenges from the perspectives of systems, humans, and cybernetics. Fortunately, there is a significant emergence of technologies including monitoring and active identification, wired and wireless communications, and intelligent agent systems. Brought together effectively these technologies could complement each other and be integrated to provide unique solutions to meet the needs of decision-makers.

In the past decade a number of international groups have been formed to develop collaborative agent, holonic, and multi-agent systems. Within the Intelligent Manufacturing Systems Program, for example, there have been several such projects. However, many of these initiatives have already concluded and others are coming to conclusion.

This presentation will describe a new initiative of the IEEE SMC Society that consists of four core competency areas and four horizontal market application areas as follows:

Competency Areas Application Areas

We will provide examples of the research challenges in the core competency and application areas from manufacturing and distributed energy systems. Because agent communications systems requires a distributed and scalable hardware infrastructure, we will also describe our research in distributed wireless 802.11 communication that provides multi-hop, peer to peer connectivity with adaptive routing and priority scheduling. Finally, the presentation will identify opportunities for using soft computing techniques and related technologies in our research projects.

We shall encourage dialogue and discussions among the participants to explore opportunities for collaboration.

Applications Variable and Feature Selection in Bioinformatics

Isabelle Guyon

Variable and feature selection have become the focus of much research in areas of application for which data sets with tens or hundreds of thousands of variables are available, particularly in text processing and medical diagnosis with genomics and proteomics data. The problem consists in removing the input variables that are either not informative or redundant. The objectives include improving the performance of predictors, providing faster and more cost-effective predictors, and providing a better understanding of the underlying process that generated the data.

We will present a historical perspective and outline the advantages and limitations of various approaches. We will show small examples that illustrate particular difficulties of the problem. Our examples will include DNA microarray analysis and protein profiling with mass spectrometry. We will explain the design of the NIPS2003 feature selection challenge and analyze the results of the benchmark.

Evolution of the Laws that Deal with the Utilization of Information Networks

Babak Hodjat, Adam Cheyer (SRI), and Russ Berg

Today, three Laws are used to explain how the value of a network, comprised of computational power and informational content, increases as the network expands: Moore's Law, Metcalf's Law, and Reed's Law. In order to achieve the exponential benefits promised by the Three Laws, nodes in the value network must possess a shared communication/information model, and there is a cost to acquiring this knowledge. In this presentation a new perspective is proposed that explores this effect, evolves the above three laws to take into account a knowledge-acquisition cost constant C, and concludes that technologies that bring this constant as close to zero as possible in real practice are the key to unlocking the exponential value inherent in information networks.

A Decentralization Method For Modeling The Multiple Levels Of Organization And Function Within The Liver

C. Anthony Hunt and Glen E. P. Ropella

Development of a biological model is driven by the experimental context in which it will be used. Hence, computer models are often overfitted to a single, unique, experimental context and fail to be useful in other situations. So doing severely limits the model's usefulness, effectively blocking inferential extensions to somewhat different conditions such as a hypothesized new treatment intervention. To solve this problem, multiple, separate models of a biological system at different levels of organization are required to understand and adequately represent their behaviors. In this poster, we present the basics of a new modeling method, FURM (Functional Unit Representation Method) that attempts to address this problem. Here we focus on the primary functional unit of the liver within an in silico isolated perfused rat liver (IS-IPRL). FURM [http://biosystems.ucsf.edu/Researc/furm/index.html] decouples the various aspects of functional units. It uses a middle-out model design strategy that enables and encourages selection of different models. It is an example of a new class of generative biological simulation models whose components are easily joined and disconnected and are replaceable and reusable. It works by decentralizing the modeling process without requiring that all of the data be of a specific type. FURM does not require any particular formalism. Rather the experimental framework is formulated using Partially Ordered Sets. We follow four fundamental guidelines: 1) standardize interfaces to multiparadigm, multimode, and trans-domain models, 2) use discrete interactions, 3) enable knowledge discovery by designing for an extended model life cycle, and 4) define observables that will submit to a similarity measure. A data model represents the biological system. An established in vitro liver perfusion protocol is the source of our experimental data in [J Pharmacokin Biopharm27:343-82, 1999]. Two in silico system models are implemented. RefModel is the accepted, reference mathematical model [JPET297:780-789, 2001]. ArtModel is our functional unit model that assumes that liver function, as a whole, is an aggregate of lobule function, that sinusoids are primarily vascular objects, and that transit time for perfusate is governed by stochastic interactions between various agents inside the vascular structures in combination with the perfusion pressure at the inlet catheter and lobule portal vein. The IS-IPRL strives to replicate the experimental procedure that has provided the experimental data. The four primary assumptions are: (1) Outflow profiles alone are lossy projections of liver behavior. Thus, physiologically accurate models are necessary to begin fully exploring the liver behavior space. (2) Hepatic vascular structure and the arrangement of lobules within a lobe can be represented by a directed graph. (3) The primary functional unit is the lobule. (4) Outflow for sucrose, but not metabolized or transported solutes, is solely a function of the extracellular (vascular cavity) space and its geometry. Within lobules agents representing sinusoidal segments (SS) are located at each graph node. Agents within each SS represent functionalities within cellular and subcellular spaces. The explicit hypothesis being tested is that the selected parameter vectors cause the model to generate output that is experimentally indistinguishable from that seen in the in vitro data. A similarity measure is used to automate the evaluation of the solution sets put forth by the models. Those results make possible automatic searches of the parameter space for regions that solve the problem by matching the in vitro data. To date different parameterizations of a mathematical model have been needed to account for outflow data for two different solutes. One IS-IPRL model now accounts for the hepatic outflow profiles of both sucrose and diltiazem and is being extended to account simultaneously for five additional drugs and to be able to shift to represent either of two disease states.

Applications of Soft Computing: Control, Robotics, Simulation and Image Processing

Mo Jamshidi

One of the biggest challenges of any control paradigm is being able to handle complex systems under unforeseen uncertainties. A system may be called complex here if its dimension (order) is too high and its model (if available) is nonlinear, interconnected, and information on the system is uncertain such that classical techniques cannot easily handle the problem. Knowledge about such systems is a key attribute which is not often exploited for design and synthesis. Soft computing, a consortium of fuzzy logic, neuro-computing, genetic algorithms and genetic programming, has proven to be powerful tools for analysis and design of many complex systems. For such systems the size soft computing control architecture will be nearly infinite. Examples of complex systems are power networks, space robotic colonies, national air traffic control system, an integrated manufacturing plant, Hubble Telescope, a swarm of robotic agents, etc. In this talk a survey of many soft computing based applications in control, robotics, simulation and image processing will be given.

Some Opportunities for Soft Computing Applications

Mo Jamshidi

In this lecture we present a series of what the speaker considers excellent application opportunities for soft computing tools with national and international implications. Economy and security are two of the most important issue of not just today, but also for tomorrow and years to come. These opportunities are intelligent multi-agent systems (e.g. V-Lab®, presented the day before) and autonomous agents monitoring, recognizance and border security. Another opportunity, not unrelated to national security, is the application of soft computing for energy efficiency of the so-called “Industries of the Future”, where SC tools can be used to save energy, reduce waste or can help avoid production of defective products, reduce energy requirements for automation and remote operations, increase efficiency of existing processes via sensing and information technology systems where demand and automated production modeling can be integrated, etc. Another opportunity exists in the chemical process systems, where relatively little effort has been made to use SC tools. Such problems as water treatment plants, oil extraction, etc. are candidate applications. Finally, two other opportunities are diagnostics/prognostics of national defense systems and image processing applications and remote sensing for the study of earth.

The Role of User Modelling in Information Retrieval from the WWW

Robert I. John

Information Retrieval from the WWW through the use of search engines is known to be unable to capture effectively the information needs of users. The approach taken in our work is to add intelligence to information retrieval from the WWW, by the modelling of users to improve the interaction between the user and information retrieval systems. In other words, to improve the performance of the user in retrieving information from the information source. To affect such an improvement, it is necessary that any retrieval system can somehow make inferences concerning the information the user might want. The system then can aid the user, for instance by giving suggestions or by adapting any query based on predictions furnished by the model.

We have taken two separate approaches. Firstly, by a combination of user modelling and fuzzy logic a prototype system has been developed (the Fuzzy Modelling Query Assistant (FMQA)) which modifies a user’s query based on a fuzzy user model. The FMQA was tested via a user study which clearly indicated that, for the limited domain chosen, the modified queries are better than those that are left unmodified.

As part of a European project, ELVIL, we have developed a Virtual Librarian. The European Legislative Virtual Library (ELVIL) is an Internet-based portal of information sources on European law and politics (http://elvil.sub.su.se) which provides software gateways with a uniform WWW-interface to national and European parliamentary databases, a searchable index to other sources of legal and political information on the WWW, and collections of learning resources on European law and politics. Information Retrieval from the WWW through the use of search engines is known to be very ineffective at meeting information needs of users. The approach taken in this work is to add intelligence to information retrieval from the WWW and online databases, by the modelling of users, using fuzzy logic, to improve the interaction between the user and the information retrieval system ELVIL. By a combination of user modelling and fuzzy logic a system has been developed (the Virtual Librarian (VL)) which takes a fuzzy user model and search information to choose the most appropriate databases or web sites within the ELVIL portal.

Knowledge Based Neurocomputing for Data Analysis and Knowledge Discovery in Bioinformatics

Nikola Kasabov

Problems in bioinformatics and medical decision support are defined and solved with the use of various soft-computing methods and results are compared. The problems include: gene expression data analysis for cancer and other disease profiling; promoter recognition; modeling gene regulatory networks; integrating gene and clinical data; cardio-vascular risk prognosis; renal function evaluation; survival prognosis; gene regulatory networks for brain function analysis. There are five main phases of information processing and problem solving in most bioinformatics systems that are covered for each of the above problems, namely: data pre-processing and filtering; feature evaluation and feature selection; model creation and evaluation; knowledge extraction; adaptation to new data.

NeuCom: A Neuro Computing Environment for Evolving Intelligence

Nikola Kasabov

A comprehensive software environment NeuCom (www.theneucom.com) for data mining, modeling and discovery will be demonstrated and will be made available to participants. NeuCom includes various methods for data visualization, feature selection, data analysis, modeling, rule extraction and adaptation. As a special method for adaptive learning of data and knowledge integration, evolving connectionist systems (ECOS) are included [1]. ECOS are multi-modular connectionist architectures that facilitate modeling of evolving processes and knowledge discovery. It is a collection of neural networks that operate continuously in time and adapt their structure and functionality through a continuous interaction with the environment and with other systems. As many processes in biology are dynamically evolving, their modeling requires evolving methods and systems, i.e. requires evolving intelligence. In the area of bioinformatics for example, new data is being made available with a tremendous speed that would also require models to be continuously adaptive. The talk adds to the traditional statistical and artificial intelligence methods the new methods of evolving systems, so that participants can compare the various approaches and apply them to their problems.

[1] N.Kasabov, Evolving Connectionist Systems: Methods and Applications in Bioinformatics, Brain Study and Intelligent Machines, Springer Verlag, 2002 (www.springer.de)

Soft Computing Tools for Gene Similarity Measures in Bioinformatics

James M. Keller and Joyce Mitchell

One of the most important objects in bioinformatics is a gene product (a protein or an RNA). In clustering and subsequent knowledge discovery on unknown gene products, the principal features are the gene sequence and expression values found following a microarray experiment. The question arises as to what is the function of this gene product and is it similar in function or structure to other up-regulated or down-regulated gene products. Many (dis)similarity measures have been proposed to measure closeness of sequences. However, for many gene products, additional functional information comes from the set of Gene Ontology (GO) annotations and the set of journal abstracts related to the gene product. For these genes, it is reasonable to include similarity measures based on the terms found in the GO and/or the index term sets of the related documents (MeSH annotations). In both cases we deal with comparing two sets of terms arranged in a taxonomy (GO or MeSH.). Some measures have been constructed to assess closeness of terms in a taxonomy, including shortest path length between terms and information theory-related values where node probabilities are estimated using a corpus of relevant documents. Utilizing such factors in addition to sequence and expression should aid in the process of knowledge discovery. It will be easier to annotate clusters, for example, when they share common descriptive terms. When an unknown gene product joins the group via sequence and expression, it is reasonable to conjecture that this gene will also share the cluster annotations (at least partially).

In this paper we propose a fuzzy measure-based similarity (FMS) for computing the similarity of two sets of terms found in a taxonomy (and hence, the two gene products annotated with terms from the taxonomy). The advantage of FMS is that it takes into consideration the context of the whole set when computing the similarity. For the case when the objects are annotated to an ontology, we propose a method of dealing with the case, since it is likely that two gene products will not be described by identical ontology terms. In addition, we propose a modification of the Resnick similarity measure such that the similarity between two ontology terms produces a number in [0,1]. In dealing with large groups of documents describing the objects under consideration, not only do we determine the similarity between the document pairs, but, by introducing the Choquet integral to the scenario, we can fuse this partial agreement function on pairs of documents into a single value relating the gene products. The measures for the final integral fusion can be tailored to produce order weighted average (OWA) operators (e.g., at least two documents must support the connection) or can be based on assessments of the "worth"; of individual and subsets of documents towards building the strength of connection. We present examples of FMS calculation for specific situations where two genes are described by a set of terms from the Gene Ontology, for two abstracts related to those genes, and for multiple abstract fusion. We also compare our measures to others from the literature.

A Proposed Fuzzy Optimal Control Model to Minimize Target Tracking Error in a Dynamic Hedging Problem with a Multi-Asset, Best-of Option

Mohammad Khoshnevisan and Sukanto Bhattacharya

The objective of our present paper is to derive a fuzzy optimal control model to explicitly derive the optimal rebalancing weights (i.e. dynamic hedge ratios) to engineer a structured financial product out of a multi-asset best-of option. The target function is actually the total cost of hedging (tracking) error over the investment horizon t = 0 to T, which is assumed to be partly dependent on the governing Markovian process underlying the individual asset returns and partly on randomness i.e. pure white noise. To derive the necessary conditions for the fuzzy optimal control, we consider the problem: Minimize \int_{0}^{T} f0(r, v, t) dt subject to the condition drj/dt = fj (r, v, t), where r is the vector of n state variables i.e. the vector of returns on the n assets underlying the best-of option and v is the vector of m fuzzy control variables i.e. the values of the replicating portfolio for m different choices of the dynamic hedge ratios. As the choice set is fuzzy, the fuzzy control vector is a subnormal fuzzy subset [p1/v1, p2/v2 ... pn/vn]; pj is the membership grade of the jth portfolio at time t.

Fuzzy Logic in the Modelling of Structures and Functions of Protein Sequences

Victor Korotkikh

A major aim of bioinformatics is to contribute to the understanding of the interrelationships between protein sequence, structure and function. In this context approaches that can suggest structures and processes associated with sequences may be of special interest. Moreover, if such a structure and a process turn to work consistently with observations then it is become even possible to speculate about them as laws of the phenomena. It appears then naturally to ask why the structure and the process are special to the extent that they are able to describe the nature of living things. In particular, whether they are fundamental or it is possible to explain them in terms of deeper structures and processes. This may explain why irreducible structures and processes would come to play a special role in the modeling. Indeed, there can be no further reductions possible in their instances only.

In the paper we present an approach within which sequences become associated with a hierarchical structure that has integer relations as the elements. The ultimate building blocks of the structure are the integers. The other elements are made of them by an organizing principle. The structure is irreducible in the sense that its existence is based on the integers only and the other elements are made in the full control of arithmetic. This reduces the explanations about the elements, and why it is they are made the way they are, to explanations that reveal why it is integers exist and arithmetic holds the way it does.

A new type of process can be revealed as the result of an interpretation of the relationships existing between the elements of the structure. In particular, the relationships between the integer relations are specified the organizing principle, which describes how an element of one level of the structure can be produced from elements of the previous one. Notably, this may be interpreted as a formation process. The structure has a natural ability to integrate processes. This may be helpful to model how the function of protein sequences as a new whole emerges from the functions of these sequences as separate entities.

In general, the approach is contrasting, because it relies on structures, with integer relations as the elements, in order to model structures and functions of protein sequences. Although the structures and the processes are rigorously defined mathematically, in computations however they resist to be easily described and processed. In the talk we discuss that fuzzy logic can propose an efficient solution to this problem.

Fuzzy Spectral Patterns for Information Integration in a Search Engine

Galina Korotkikh

Recently, Lotfi Zadeh suggested that the development of a search engine with the ability to synthesize an answer to a query from different information elements would be required. In the search the information elements are specified by the query and their integration into an aggregated information element serves to provide a basis for the answer.

In general, the problem of information integration is very challenging and remains unresolved. In the Internet search context it may be characterized by a huge number of information elements involved. Therefore, an identification and understanding of possible patterns related to the information integration may turn to be useful in the development of a search engine.

In the paper we study this problem in terms of the eigenvalue spectra of the variance-covariance matrices. They are applied to binary sequences as information elements to model the information integration. The eigenvalue spectrum of the variance-covariance matrix is used to investigate the result of the information integration. It is suggested that some patterns of the eigenvalue spectra may be helpful in the understanding of the information integration.

We demonstrate that there exists a certain pattern dynamics of the eigenvalues within which the spectra of the variance-covariance matrices may be described. We reveal a mechanism that underlies the eigenvalue dynamics and show that its functioning can find efficient computational expression by a quadratic trace of the variance-covariance matrix. It is proposed how the pattern of the eigenvalue spectra, specified by the dynamics and the mechanism, may be interpreted to help us in the understanding of the information integration.

It is also discussed that the fuzzyfication of the eigenvalues has played a crucial role in the identification of the dynamics and the mechanism. This may support that fuzziness is an integral part of the information integration.

Computational Intelligence and Modelling in Materials Science and Processing

John A. Meech

To a great extent the future welfare of the world depends on developing new materials and processes. In all fields of endeavour there is need for materials with new properties that support the necessary advances. These fields encompass among others: robotics, computer hardware/software, automobiles and transportation, communication systems, medicine and biological systems, environmental protection, energy production and distribution, aircraft and aerospace, rapid transit systems and railroads, undersea exploration, building construction and other structures, missiles and weapon systems, commercial "white" products, etc. As each of these fields evolve into new areas of science and technology, demands are made for increased speed, increased productivity or bandwidth, increased storage capacities, increased strength or mechanical resistance, increased efficiency or effectiveness, adaptability to increased complexity, decreased size, decreased pollution levels and emissions, decreased costs, decreased cycle times, integration of existing processes, and doing this while encorporating "intelligence" within specific materials, processes or products.

This paper reviews some recent advances in the application of "intelligent" methods in the production of new materials and manufacture of new products. In some cases, the material embodies its "SMARTness" into an ability to self-adapt or respond to environmental changes. In other cases, the intellect is embodied into the processes used to derive the materials - primary processing, extraction, finishing, assembling, and delivery. Finally the use of "intelligence" to perform or support simulation models of new materials and/or their processes is presented. Both empirical data-derived models as well as First Principles modeling are discussed.

Adaptive Clustering in Medical Image Segmentation as a Diagnostic Tool

Sunanda Mitra, Shuyu Yang

Image analysis techniques have been broadly used in computer-aided medical analysis and diagnosis in recent years. Computer aided image analysis is an increasingly popular tool in medical research and practice, especially with the increase of medical images in modality, amount, size and dimension. Image segmentation, a process that aims at identifying and separating regions of interests from an image, is crucial in many medical applications such as localizing pathological regions, providing objective quantitative assessment and monitoring of the onset and progression of the diseases, as well as analysis of anatomical structures. For clinical applications of segmentation, a compromise between the accuracy and computational speed of segmentation techniques is needed. Optimal segmentation processes based on statistical and adaptive approaches and their applicability to clinical settings will be addressed using diverse modalities of images. Current drawbacks of automated segmentation methodologies stem mostly from non-uniform illumination, inhomogeneous structures, and the presence of noise in acquired images. The effect of preprocessing on the accuracy of segmentation will be discussed. The superior performance of advanced clustering algorithms based on statistical and adaptive approaches over traditional algorithms in medical image segmentation will be presented.

Causality Recognition For Data Mining In An Inherently Ill Defined World

Lawrence J. Mazlack

Common sense causal reasoning occupies a central position in human reason­ing. It plays an essential role in hu­man decision-making. Considerable effort has been spent examining causation. Philosophers, mathematicians, computer scientists, cognitive scientists, psychologists, and others have formally explored questions of causation beginning at least three thousand years ago with the Greeks.

Whether causality can be recognized at all has long been a theoretical speculation of scientists and philosophers. At the same time, in our daily lives, we operate on the commonsense belief that causality exists.

Causal relationships exist in the commonsense world. If an automobile fails to stop at a red light and there is an accident, it can be said that the failure to stop was the accident's cause. However, conversely, failing to stop at a red light is not a certain cause of a fatal accident; sometimes no accident of any kind occurs. So, it can be said that knowledge of some causal effects is imprecise. Perhaps, complete knowledge of all possible factors might lead to a crisp description of whether a causal effect will occur. However, in our commonsense world, it is unlikely that all possible factors can be known. What is needed is a method to model imprecise causal models.

Another way to think of causal relationships is counterfactually. For example, if a driver dies in an accident, it might be said that had the accident not oc­curred; they would still be alive.

Our common sense understanding of the world tells us that we have to deal with imprecision, uncertainty and im­perfect knowledge. This is also the case of our scientific knowledge of the world. Clearly, we need an algorithmic way of handling imprecision if we are to computationally handle causality. Models are needed to algorithmically consider causes. These models may be symbolic or graphic. A difficulty is striking a good balance between precise formalism and commonsense imprecise reality.

An Evolutionary Computing Approach to Minimize Dynamic Hedging Error

Saeid Nahavandi and Mohammad Khoshnevisan

The objective of our present paper is to derive a computationally efficient genetic pattern learning algorithm to evolutionarily derive the optimal rebalancing weights (i.e. dynamic hedge ratios) to engineer a structured financial product out of a multi-asset, best-of option. The stochastic target function is formulated as an expected squared cost of hedging (tracking) error which is assumed to be partly dependent on the governing Markovian process underlying the individual asset returns and partly on randomness i.e. pure white noise. A simple haploid genetic algorithm is advanced as an alternative numerical scheme, which is deemed to be computationally more efficient than numerically deriving an explicit solution to the formulated optimization model. An extension to our proposed scheme is suggested by means of adapting the Genetic Algorithm parameters based on fuzzy logic controllers.

web user profiling and web personalization based on data mining and soft computing

Olfa Nasraoui

Intelligent Web personalization aims at adapting a user's interaction with the Web information space based on information gathered about the user. A complete automated Web personalization system is generally based on Web usage mining to discover useful knowledge about user access patterns, followed by a recommendation system to act on this knowledge in order to respond to the users' individual interest, in a manner transparent to the user, and while protecting the user's privacy and anonymity.

Web usage mining has recently attracted attention as a viable framework for extracting useful access pattern information, such as user profiles, from massive amounts of Web log data for the purpose of Web site personalization and organization. These efforts have relied mainly on clustering or association rule discovery as the enabling data mining technologies. Typically, data mining has to be completely re-applied periodically and offline on newly generated Web server logs in order to keep the discovered knowledge up to date. Also, to date, the clustering techniques that have been used for Web usage mining have rather scored low on scalability both in terms of time and memory requirements, and simply cannot be expected to keep up with the huge flux of Web clickstream data on today's busy websites. In addition to their lack of scalability and difficulty to adapt in the face of continuously evolving patterns, current clustering techniques, such as most KMeans variants, also suffer from one or more of the following limitations: requirement of the specification of the correct number of clusters/profiles in advance, sensitivity to initialization, sensitivity to the presence of noise and outliers in the data. Other techniques relying on association rule discovery suffer from their sensitivity to various thresholds such as support and confidence, as well as sparsity of the data. Hence, there is a crucial need for scalable, noise insensitive, initialization independent techniques that can continuously discover possibly changing/evolving Web user profiles without any stoppages or reconfigurations.

We overview some recent evolutionary computation techniques for Web mining. Then, we present a new scalable clustering methodology that gleams inspiration from the natural immune system to be able to continuously learn and adapt to new incoming patterns. The Web server plays the role of the human body, and the incoming requests play the role of foreign antigens/bacteria/viruses that need to be detected/recognized by the proposed immune based clustering technique. Hence, our clustering algorithm plays the role of the cognitive agent of an artificial immune system, whose goal is to continuously perform an intelligent organization of the incoming noisy data into clusters/patterns. Our approach exhibits superior learning abilities, while at the same time, requiring modest memory and computational costs. Like the natural immune system, the strongest advantage of immune based learning compared to current approaches is expected to be its ease of adaptation to the dynamic environment that characterizes several applications, particularly in mining data streams. We illustrate the ability of the proposed approach in detecting clusters in noisy data sets, and in mining user profiles from Web clickstream data in a single pass under different usage trend sequence scenarios.

Finally, we present several soft computing approaches to high performance mass profile-based web recommender engines that we have recently developed.


This work is supported by a National Science Foundation CAREER Award (NSF-IIS-0133948) to O. Nasraoui.

Context-directed information access

Charles Ortiz

I will present work that represents a departure from the master-slave metaphor of human-system interaction --- in which a user first (precisely) communicates a request that is then followed by a response --- with one based on collaboration in which a user and an application work jointly toward some goal. Such collaboration leads to a somewhat different view of communication, particularly in information-gathering tasks such as those prevalent on the web. In such tasks, the fruits of collaborations take the form of the establishment of what we call an access perspective. An access perspective reflects the inherently multifaceted way in information at one location may be accessed from another, depending on the intended use of that information and the required level of detail.  Access perspectives are dynamic entities that are informed by the structure of a user's activity. Conventional, nonadaptive user interfaces characteristically provide a static access method.  I will describe methods for representing a context tree which establishes a particular access perspective; as a user explores a particular branch of a context tree, the context becomes more and more constrained and the access perspective becomes, correspondingly, narrower.

Graph Data Management for Biology

Franz Olken

In this talk we shall discuss the use of graph data management for a variety of biological applications. Simple graphs are composed of a set of nodes and a binary relation comprising the edges which connect pairs of nodes.

We shall emphasize applications related to the representation and querying of biopathways databases, e.g., metabolic pathways, signal transduction pathways, and genetic regulatory networks. Other potential applications of graph data management to biology include: chemical structure graphs, protein interaction networks, phylogenetic trees, taxonomies of chemicals, proteins, enzymes and diseases, partonomies (e.g., in anatomy), topological adjacency relations (in image analysis), contact graphs (for 3D protein structure), bibliographic citation graphs, food webs, biogeochemical cycles, gene clusterings, partial order graphs for DNA multiple sequence alignments, genetic maps, operon and regulon structures, sequence overlap graphs for shotgun DNA sequencing, database schemas and mappings among schemas, data provenance (lineage), hypertext, semantic web applications, laboratory protocols, etc.

Graph data models for biology come in a number of variants: undirected and directed graphs, simple graphs, nested graphs, multigraphs, and hypergraphs. We will mention these variants and illustrate their applications.

Graph data management offers two major advantages for biopathways applications: naturalness of representation of pathway data, and ease of querying pathway data. It is the latter issue which is more important.

Graph data management systems permit users to frame queries in terms of graph operations, e.g., subgraph isomorphism, shortest paths, etc. which would be difficult to express or compute in conventional (relational) DBMS systems. We discuss a number of graph queries in the talk, e.g., subgraph homomorphism queries.

Graph data management systems typically treat individual fragments of the database more homogeneously (as either nodes or edges) than relational databases which partition the database into many specialized relations. In a GDMS the analog of relation structure is encoded as edges which indicate types of nodes. While the relational storage structures offer advantages in performance on fixed structure queries, the homogeneous graph data model is much easier to use in posing queries which allow paths to span many different possible relations (node types). It is this storage homogeneity which facilitates pattern matching and path queries in graph databases. In contrast, similar queries in relational DBMS involve large numbers of union queries over the various possible relations which might participate in a path or subgraph pattern match. We will discuss this issue and (briefly) some related comparisons to logic and object oriented database management systems.

We will illustrate the talk with references to some major biopathways databases. Time permitting we will also mention the role of RDF (Resource Description Framework) as a a graph data model. We will conclude with a brief survey of some alternative approaches to implementation of graph database management systems.

This is joint work with Kevin D. Keck and Vijaya Natarajan (both at LBNL). Further information on the Biopathways Graph Data Manager Project may be found at http://www.lbl.gov/~olken/graphdm/graphdm.htm The work is funded by DARPA Biocomp Program (via the Biospice Project at LBNL (PI: A. Arkin)), and DOE Genomes to Life Program (via VIMSS GTL project at LBNL (PIs: A. Arkin and T. Hazen) and Synechococcus GTL project at Sandia National Lab (PI: G. Heffelfinger, LBNL PI: A. Shoshani). Submitted by Frank Olken on 2003-12-08.

Machine Intelligence, Granular Computing and Data Mining

Sankar K. Pal

Emergence of data mining and knowledge discovery from pattern recognition point of view is illustrated. Significance of integrating various soft computing tools for efficient learning is described. Role of granular computing in data mining is given more emphasis. Two examples, demonstrating integrations of fuzzy sets, artificial neural networks, genetic algorithms and rough sets for efficient classification, rule generation and rule evaluation, and for granular case generation in case based reasoning problems, are provided along with their application specific merits on real life data. The significance of rough-fuzzy granulation for both reducing the computation time and improving the performance is explained. The talk concludes explaining the relation of rough-fuzzy case generation with the recently emerged computational theory of perception (CTP) and their applications in Web mining problems.

Fuzzy Logic, Neutrosophic Logic, and Applications

Florentin Smarandache and Mohammad Khoshnevisan

A Small Survey of the Evolution of Logics from Boolean to fuzzy and neutrosophic is presented. Afterwards the neutrosophic logic components are introduced followed by the definition of neutrosophic logic (NL), based on non-standard analysis, and neutrosophic logic connectors which are based on set operations.

Neutrosophic Logic Components:

Let T, I, F be standard or non-standard real subsets of ] -0, 1+ [,

with sup T = t_sup, inf T = t_inf,
sup I = i_sup, inf I = i_inf,
sup F = fsup, inf F = finf,
and nsup = tsup+isup+fsup,
ninf = tinf+iinf+finf.

The sets T, I, F are not necessarily intervals, but may be any real sub-unitary subsets: discrete or continuous; single-element, finite, or (countably or uncountably) infinite; union or intersection of various subsets; etc. They may also overlap. The real subsets could represent the relative errors in determining t, i, f (in the case when the subsets T, I, F are reduced to points). In the paper, T, I, F, called neutrosophic components, will represent the truth value, indeterminacy value, and falsehood value respectively referring to neutrosophy, neutrosophic logic, neutrosophic set, neutrosophic probability, neutrosophic statistics. This representation is closer to the human mind reasoning. It characterizes/catches the imprecision of knowledge or linguistic inexactitude received by various observers (that's why T, I, F are subsets - not necessarily single-elements), uncertainty due to incomplete knowledge or acquisition errors or stochasticity (that's why the subset I exists), and vagueness due to lack of clear contours or limits (that's why T, I, F are subsets and I exists; in particular for the appurtenance to the neutrosophic sets). One has to specify the superior (x_sup) and inferior (x_inf) limits of the subsets because in many problems arises the necessity to compute them.

Definition of Neutrosophic Logic: A logic in which each proposition is estimated to have the percentage of truth in a subset T, the percentage of indeterminacy in a subset I, and the percentage of falsity in a subset F, where T, I, F are defined above, is called Neutrosophic Logic. We use a subset of truth (or indeterminacy, or falsity), instead of a number only, because in many cases we are not able to exactly determine the percentages of truth and of falsity but to approximate them: for example a proposition is between 0.3-0.4 true and between 0.6-0.7 false, even more imprecise: between 0.3-0.4 or 0.45-0.50 true (according to various analyzers), and 0.6 or between 0.66-0.70 false. The subsets are not necessary intervals, but any sets (discrete, continuous, open or closed or half-open/half-closed interval, intersections or unions of the previous sets, etc.) in accordance with the given proposition.

A subset may have one element only in special cases of this logic.

The differences between IFL and NL [and the corresponding Intuitionistic Fuzzy Set (IFS) and Neutrosophic Set (NS)] are:

a) Neutrosophic Logic can distinguish between absolute truth (truth in all possible worlds, according to Leibniz) and relative truth (truth in at least one world), because NL(absolute truth)=1+ while NL(relative truth)=1. This has application in philosophy (see the neutrosophy). That's why the unitary standard interval [0, 1] used in IFL has been extended to the unitary non-standard interval ]-0, 1+[ in NL. Similar distinctions for absolute or relative falsehood, and absolute or relative indeterminacy are allowed in NL.

b) In NL there is no restriction on T, I, F other than they are subsets of ]-0, 1+[, thus: -0[inf T + inf I + inf F[ sup T + sup I + sup F [ 3+. This non-restriction allows paraconsistent, dialetheist, and incomplete information to be characterized in NL {i.e. the sum of all three components if they are defined as points, or sum of superior limits of all three components if they are defined as subsets can be >1 (for paraconsistent information coming from different sources) or < 1 for incomplete information}, while that information can not be described in IFL because in IFL the components T (truth), I (indeterminacy), F (falsehood) are restricted either to t+i+f=1 or to t2 + f2 [1, if T, I, F are all reduced to the points t, i, f respectively, or to sup T + sup I + sup F = 1 if T, I, F are subsets of [0, 1].

c) In NL the components T, I, F can also be non-standard subsets included in the unitary non-standard interval ]-0, 1+[, not only standard subsets included in the unitary standard interval [0, 1] as in IFL.

d) NL, like dialetheism, can describe paradoxes, NL(paradox) = (1, I, 1), while IFL can not describe a paradox because the sum of components should be 1 in IFL.

Applications of Fuzzy and Neutrosophic Logics: Further one gives examples of fuzzy and neutrosophic logics used in the reconciliation between theoretical and market prices of long-term options contracts, in extension of the MASS model as a cost-optimal relative allocation of facilities technique by the incorporation of neutrosophic statistics and the DSm (Dezert-Smarandache) combination rule, and in conditional probability of actually detecting a financial fraud – a neutrosophic extension to the application of Benford's first-digit law.

Original work consists in the definition of neutrosophic logic (NL) and neutrosophic logical connectors as an extension of intuitionistic fuzzy logic (IFL) and the comparison among NL and other logics, especially the IFL, and in their applications.

Basic Principles of Information Monitoring Systems

Alexander Ryjov

This article describes main ideas of Information Monitoring Systems (IMS) and applications of IMS in real-world problems. Information monitoring systems relate to a class of hierarchical fuzzy discrete dynamic systems. The theoretical base of such class of systems is made by the fuzzy sets theory, discrete mathematics, and methods of the analysis of hierarchies which was developed independently in works of Zadeh, Messarovich, Saaty and others. IMS address to process uniformly diverse, multi-level, fragmentary, unreliable, and varying in time information about some problem/process. Based on this type of information IMS allow perform monitoring of the problem's/process' evolution and work out strategic plans of problem/process development. These capabilities open a broad area of applications in business (marketing, management, strategic planning), socio-political problems (elections, control of bilateral and multilateral agreements, terrorism), etc. One of such applications is a system for monitoring and evaluation of state's nuclear activities (department of safeguards, IAEA) has been shortly described in the report.

Fuzzy temporal predicates in data mining

Thomas Sudkamp

The structure and precision of the data affects the type of relationships that can be analyzed by knowledge discovery and data mining algorithms. Data items are frequently sets, tuples in relational databases, or elements from a temporal stream. The former two representations consider individual data items as independent while temporality induces an ordering on the data. When data items are independent, data mining algorithms seek to discover frequently recurring associations within individual items. When distinct data items are related, the task of knowledge discovery is to identify relationships between the items.

In this presentation we review the types of relationships produced by standard data mining techniques. In particular we are interested in the representation of imprecision in both the data and the resulting associations. For independent data, fuzzy sets have been used to partition attribute domains to extend quantitative association rules to imprecise categories. Temporal relationships consider data from one or multiple sources in which the sole link between the data may be the time value. Association rules for temporal data may include imprecise durations and temporal constraints. Temporal durations specify the time an event is required to continue to satisfy a proposition, for example event A occurs for "a long period" of time. A constraint specifies the temporal relationship between the occurrence of distinct events; B occurs shortly after A. We will discuss the constraints, data representations, and search strategy needed for analyzing fuzzy temporal associations in data from multiple sources.

Mining opinion

James G. Shanahan

Today, much of product feedback is provided by customers/critiques online through websites, discussion boards, mailing lists, and blogs. When trying to make strategic decisions (a product launch, a purchase), using a websearch will return many useful but heterogeneous and, increasingly, multilingual opinions on a product. Generally, the user will find it very difficult and time consuming to assimilate all available information and make an informed decision. To date most work in automating this process has focused on a monolingual setting. Existing monolingual approaches will overviewed. In addition, I will describe our preliminary work on mining product ratings in a multilingual setting. The proposed approaches are automatic, using a combination of techniques from classification and translation, thereby alleviating human-intensive construction and maintenance of linguistic resources.

Semantics Powered Bioinformatics: Semantic Search, Processes and Analytics

Amit Sheth

Bioinformatics is about providing biologists and those in allied disciplines with ability to exploit information - information that is increasingly distributed, heterogeneous and massive. Bioinformatics has already succeeded in utilizing database management, workflow and information retrieval technologies, which have provided syntactic search, heterogeneous data access and sharing, and limited forms of integration. A good amount of effort has also involved use of statistical and syntactic techniques to support the essential tasks of finding patterns, similarities, and matches to identify building block structures. Looking to the future, we can realize more exciting potential of bioinformatics if we have more automated ways for analysis leading to insight and discovery - to understand cellular components, molecular functions and biological processes, and more importantly complex interactions and interdependencies between them. And while lot of effort in the last decade focused on genes, next set of challenges involve more complex structures of protein and carbohydrates.

This talk focuses on semantics enabled bioinformatics. We outline increasing use of semantic techniques in bioinformatics for search, browsing, integration, analytics and discovery. Ontologies provide underpinning for most of today's semantic techniques and the Semantic Web research, and bioinformatics is one of the most aggressive adopters of ontologies among science and industry domains. We provide examples of use of multiple ontologies and Semantic Web Processes to investigate more automation in discovery and analysis. We will also weave a brief overview of the research and exciting commercial state of the art in the semantic technology (specifically ontology driven information systems).

Sample background/related material:

Computing and Current Trends in Supercomputing

Horst D. Simon

Supercomputing today in 2003/2004 is no longer based on custom built expensive vector supercomputers which were associated with the field more than a decade ago. Many of the high-end computers today are built from inexpensive commodity parts, using open source software. Consequently the application of this high-end computing technology to floating-point intensive applications in science and engineering has become a dynamic and vibrant field. Clusters and related technology are used everywhere from science and engineering departments in universities, to commercial applications in banking, telecoms, and biotech. Surprisingly the field of soft computing as represented by the topics of this workshop, has not been a strong participant in this revolution of high end computing. With some notable exceptions such as Google's search engine that runs on commodity clusters, parallel supercomputing and soft computing seem to be living in parallel universes.

In this talk, I will first survey the state of the field in supercomputing, and analyze current developments based on data from the TOP500 list. I will demonstrate with a few examples from NERSC's users community, how powerful compute resources have transformed science, and how computational science has become accepted as the third leg of science in many disciplines. Then I will offer an analysis, why these developments have had little impact on computational problems in soft computing. Based on this analysis, I will conclude with some suggestions on how both fields could interact and learn from each other in the future.

Research and Applications of Soft Computing at BT Exact

Marcus Thint

Advances in AI and Soft Computing are key to realizing the visions of the intelligent enterprise, iBusiness (intelligent e-Business), and web intelligence. BT Exact's Research and Venturing division is driven by applications and development of these key technologies for business solutions. This presentation describes projects in the Computational Intelligence Group, which focuses on software systems and tools for intelligent data analysis and information management. In particular, we discuss the use of soft computing for software agents and adaptive user profiling, data representation, information search and retrieval, document classification, text mining, and intelligent scheduling. Our goals are to increase automated reasoning, improve resource/information utilization, and assist human-machine interaction in the realm of natural language/unstructured data. The systems are targeted at both internal customers (e.g. Contact Centers), and external users.

A Framework for Effective Gene Expression Analysis and Biological Knowledge Discovery

S. M. Vincent Tseng, Lien-Ching Chen, Shih-Chiang Yang

The cDNA and Oligonucleotide microarray chips have become the most popular techniques for investigating gene expressions profiles. The high density DNA microarray technology can simultaneously monitor the expression levels of thousands of genes. How to analyze the huge experimental result and discover useful biological knowledge is an important topic. In this research we present a framework, namely GeneFilter, for effectively analyzing gene expression data and performing knowledge discovery. The main goals of the framework are as follows: 1) Design an effective model and flow for high-throughput data analysis, 2) Develop an integrated and efficient platform for gene expression analysis including data preprocessing, gene expression patterns mining and visualization modules, 3) Develop methodologies for biological knowledge discovery based on the gene expression analysis results.

The system architecture of GeneFilter is as shown in Figure 1. In the wet-Lab, biologists design the experiments for targeted diseases and conduct experiments with cDNA microarray, oligonucleotide microarray, Q-PCR, or 2D gel chip. The experimental results are then submitted to GeneFilter for the Dry-Lab analysis, including data preprocessing, statistical analysis, gene expression patterns mining, gene ranking and gene ontology analysis. Finally, the analysis results are validated by using more precise biological experiments like Q-PCR, and the validation results are feed back for refining the further analysis. In the following, we describe the main functions in each of the analysis process:

1. Data Preprocessing and Statistical Analysis

In this process, the quality of the microarray data is first examined via various statistical methods and visualization techniques like in [1]. The data of bad quality will be filtered out eventually. Then, various types of normalization methods, like Within-slide normalization, Paired-slides normalization, Multiple-slides normalization [2, 3], are provided for applications on the data. Various statistical graphs like MA Plots are also provided for examining the normalized results.

2. Gene Expression Patterns Mining

In this process, the interesting patterns hidden behind the microarray data are discovered automatically via the following data mining methods we developed based on our past experiences [7, 8, 9]: 1) Automatic and customized discovery of gene expression patterns (e.g., the up-regulated or down-regulated expression patterns), 2) Validation-based clustering for finding the nearly-optimal clustering of the gene expression in very short time, 3) Time-series analysis for discovering the activating relations between genes with tolerance of noisy data, 4) Classification analysis for modeling or contrasting gene expression patterns under different disease types (e.g., the dominating genes related to bladder cancer).

3. Gene Ranking<:p>

For the genes identified to be with specific patterns we are interested via the data mining methods, a ranking mechanism is further applied for distinguishing the possible degree of importance of these genes. The ranking mechanism takes as inputs various kinds of information and then calculates a score for each gene based on the given weights for each type of information. The input information includes the significant amount of gene expression ratio value in the microarray, the degree of relevant to the targets based on the published literatures (e.g., Gene Expression Omnibus (GEO) web site (http://www.ncbi.nlm.nih.gov/geo/ and [5]) releases the genes related to various kinds of cancers), Q-PCR results, etc.

4. Gene Ontology Analysis

For the interesting genes discovered through the above processes, three organizing networks of Gene Ontology (http://www.geneontology.org/ ) information, namely molecular function, biological process and cellular component, are provided for exploring the deeper knowledge on the genes. In GeneFilter platform, an agent program was designed for accessing the latest Gene Ontology information, and a querying system was provided for users to discover the Gene Ontology information related to specific genes

5. Analysis Validations and Cyclic Refinement

Through the above analysis process, the most interesting genes related to the analysis targets will be discovered in an automatic and efficient way. To validate the analysis results, further biological experiments should be conducted via more precise biological experiments like Q-PCR [6]. The validation results will be feed back into the analysis platform again for refining the analysis by adjusting the policies and parameters in data preprocessing, expression pattern mining and gene ranking. Through the cyclic refinement process, more and more accurate analysis results will be obtained.

In real applications, we have conducted an extensive analysis of bladder cancer microarray on GeneFilter platform for discovering the genes most influential on the progress on different stages of bladder cancer. Multiple slides of microarray with more than 10,000 genes are analyzed and genes with interesting expression patterns (e.g., up-regulated) are discovered successfully. In particular, the whole analysis process can be done within very short period due to the nice features of the platform in terms of high degree of integration and automation.


[1] S. Dudoit, Y. H. Yang, T. P. Speed, and M. J. Callow, "Statistical methods for identifying differentially expressed genes in replicated cDNA microarray experiments" in Statistica Sinica, Vol. 12, No. 1, p. 111-139, 2002.

[2] Yee Hwa Yang, Sandrine Dudoit, Percy Luu and Terry Speed, "Normalization for cDNA Microarray Data," in SPIE BiOS, San Jose, California, January 2001.

[3] X. Wang, M. J. Hessner, Y. Wu, N. Pati, S. Ghosh, "Quantitative quality control in microarray experiments and the application in data filtering, normalization and false positive rate prediction," in Bioinformatics, 19(0): 1-7, 2003.

[4] Anil K. Jain & Richard C. Dubes (1988) Algorithms for Clustering Data. Prentice Hall.

[5] Jung-Hsien Chiang and Hsu-Chun Yu, "MeKE: discovering the functions of gene products from biomedical literature via sentence alignment," Bioinformatics 2003 19: 1417-1422.

[6] M.S. Rajeevan, Vernon S.D., Taysavang N., and Unger E.R. (2001) Validation of Array-Based Gene Expression Profiles by Real-time (Kinetic) RT-PCR. J. Mol. Diagnostics 3(1):26-30.

[7] S.M. Tseng and L. J. Chen, "An Empirical Study of the Validity of Gene Expression Clustering," in Proc. 2002 Int'l Conf. on Mathematics and Engineering Techniques in Medicine and Biological Sciences (METMBS), Las Vegas, USA, 2002.

[8] S.M. Tseng and Ching-Ping Kao, "Mining and Validation of Gene Expression Patterns: An integrated approach and applications," in Informatica, vol. 27, no. 1, pp. 21-27, 2003.

[9] S. M. Tseng and Y. L. Chen, "An Effective Approach for Resolving Fundamental Problems in Time-Series Gene Expression Mining", in Proc. Int'l Workshop on Foundation and Direction in Data Mining (held with IEEE ICDM), USA, Nov., 2003.

Belief and Fuzziness: Re-structuring Epistemology

Burhan Turksen

We review briefly the basic components of the philosophical grounding of fuzzy theory from ontology to epistemology to applications. Within the context of epistemology, the notion of “belief” plays an essential role in foundationalism. In order to generalize such foundational concerns of epistemology, we propose that it be re-structured to include the language of fuzzy theory. Thus “Precisiated Natural Language, PNL” proposed by Lotfi A. Zadeh ought to be considered the sixth link in the evolutionary chain of scientific languages where meta-languages, speech, writing, mathematics, and computing languages represent the usual evolutionary links. It is well understood that all forms of language have both a communications and informatics dimension that facilitates human thought and decision making and therefore have an impact on computing with words and perceptions. I this regards, we review the connection between “belief” and “fuzzy sets” as well as classical probability and fuzzy sets.

Multi-Agent system and soft computing technologies as a fruitful symbiosis for reliable medical diagnoses

Rainer Unland

Multi-agent systems are one of the most exciting research areas in computer science at the moment. In the last years there has been a growing interest in the application of agent-based systems in health care and especially in medical diagnosis. This stems mainly from the fact that in today’s global world a fast and reliable medical diagnosis generation is of eminent importance as can be seen, for example, from the recent problems with SARS. Such highly contagious and lethal diseases can threaten the population if they are not fought immediately and with high efficiency. However, medical diagnosis can be a complex process with a lot of uncertainty and fuzziness. Thus, the integration of multi-agent system technology with soft computing technologies seems to be very promising, especially, since soft computing is tolerant of imprecision, uncertainty and partial truth. In this talk we will discuss whether and how multi-agent system technology can be used to improve and support the medical diagnosis process, especially, in situations when there is high uncertainty about the right diagnosis. Moreover, it will be shown in what way soft computing technologies can contribute to a further significant improvement of the diagnosis finding process and everything else related to it.

Emergent Information Infrastructures for Intelligent Distributed Systems : Mirroring the Universe and Life in Cyberspace

Mihaela Ulieru

We propose a fuzzy-evolutionary approach to self-organization that emulates social behavior and immunity in Cyberspace and has applications to emergency response management (Fig. 1) distributed manufacturing, medical informatics and Cybersecurities. Organized in a nested hierarchy (Fig. 2) distributed throughout the network the system consists of a hybrid mixture of static and mobile agents behaving like a Cyberorganism capable to react to unexpected changes / attacks in an optimal manner. Computational intelligence techniques endow the MAS with learning and discovery capabilities. By ‘cloning’ real-life entities into software agents, the proposed paradigm can be easily extended to the creation of emergent dynamic information infrastructures that are autonomous and proactive, capable of ensuring ubiquitous (optimal) resource discovery and allocation while at the same time self-organizing their resources to optimally accomplish the desired objectives.

Fig. 1: Emergency Scenario

Fig. 2: Holonic System

Fuzzy Logic for Genetic Networking

Paul P. Wang

This decade will probably be remembered as the “genome decade”. Almost a dozen microorganism sequences, including bacteria, and genomic sequences for man, mouse, etc. have already been completed. However, the post genome sequencing era has just begun. The focus now is on uncovering the functional organization of cells. The systemic global studies of gene expression and DNA-protein interactions in different conditions is a topic of considerable interest.

There are numerous modeling approaches of genetic networks already being proposed, but the Boolean network model discovered about three decades ago remains quite strong among researchers in the field. In this paper, the “fuzzy logic network” FLN, is proposed as a viable model for genetic networking studies. The basic properties and characteristics of the FLN are presented in this paper. The Boolean network model can be shown to be a special case of the FLN. As pointed out by numerous researchers, the Boolean network is only a coarse-grained and symbolic modeling despite the faithful modeling of the gene self-regulating process. FLN not only provides more qualitative information, but also provides some essential quantitative information about the self-regulating mechanism of the genes in the network.

One important aspect of the improvement is the possible control of the genetic network. A node may be added to a graph which represents the genetic network. Mathematically, this node represents a biological stimulus where a stimulus is any relevant physical or chemical factor which influences the network and is itself not a gene of gene product. In system and control literatures, this node is called an input function or excitation function.

Theoretically speaking, we are able to show that this control node, under certain constraints, is capable of regulating the genetic network, which is already self-regulating. To say the very least, this theoretical result may prove to be useful in experimental design. This control strategy may potentially be useful in obtaining desired “target network states”. Achieving these preferred “target network states” may have significance in cell transformation, cell repair or drug target design.

Much more remains to be explored and we hope the fuzzy logic network will be eventually be accepted as a viable model in the genomic research community. This possibility certainly exists due to FLN’s better fit with the reality in biology as well as its simplicity in computations.

Pruning, Selective Binding and Emergence of Internal Models: Applications to ICA and Analogical Reasoning

Syozo Yasui

Pruning of multi input/output neural networks is discussed and a pruning algorithm called CSDF is described. CSDF acts to induce internal models as a result of redundancy elimination and selective bindings. CSDF is used in a new ICA method based on an auto-encoder performing sensor-signal identity mapping. An internal model of the external signalmixing situation emerges due to the CSDF pruning, and the hidden units that survive the CSDF pruning reconstruct the blind source signals. This ICA method which requires no pre-processing such as whitening is characterized by its high adaptability and robustness, as is demonstrated by trouble cases such as sudden increase of the source signals, sudden failure of sensors and so on. As another example, CSDF is applied in a neural network for analogical learning/inference. Internal abstraction models together with abstraction/de-abstraction bindings are generated as a result of the CSDF structural learning coupled with the backpropagation training. The internal abstraction model acts as an attractor for new relevant dataset, a process corresponding to analogical memory retrieval.

Genetic Fuzzy Neural Agents Using Type-2 Fuzzy Reasoning for Intelligent Web Information Search Task

Yuchun Tang, Yanqing Zhang

Finding the desired information on the World Wide Web is not an easy task because the information available on the WWW is inherently unordered, distributed, and heterogeneous. As a result, the ability to search and retrieve information from the Web efficiently and effectively is a key technology for realizing the Web’s full potential

Expressing a search request is the first thing we need to solve. At almost all search scenarios, the desired information is on some Web pages. For some similar search requests, the desired Web pages share some similar content characteristics and/or structure characteristics. Traditional search methods let users to submit "keywords" that may be displayed on the desired Web pages to express their search requests. So we call them “keyword-based search” or “content-based search”. Obviously, they are not so powerful. There are 2 problems: 1 in many cases, a search request is inherently fuzzy and thus difficult or even impossible to be expressed by "crisp" keywords. Furthermore, many “partial related” Web pages can not be retrieved in a crisp way. And 2 in many, if not all, cases, some keywords provide some "structure characteristics" while others only provide "content characteristics". Unfortunately, traditional search methods do not differentiate the two kinds of keywords. In the following, “keywords” will be referred as those words that only provide content characteristics.

Retrieving information effectively is another issue we need to think about. Current search engines are known for poor accuracy: they have both low recall and low precision. Furthermore, the most relevant/desired documents are not always displayed at the top of the query result list. It means the query result is not listed in a desired order.

We proposed an intelligent web information search and retrieval model called Web Information Search Task (WIST). WIST model has two goals: one is to make the interface of a search service more expressive and another is to make information retrieval more effective. Many search requests have different content characteristics but share similar structure characteristics. These structure characteristics are expressed by simple structure rules. Basically, a structure rule is an IF-THEN formula defined on web pages’ URL, Title, Text, Input Links, Output Links, or other related sections. If the IF part is satisfied by a web page, it may be desired, otherwise it may be not desired. These structure rules will function as an input fuzzifier so we can make fuzzy reasoning to derive the “desirability”, the possibility whether a web page is desired and how much is the possibility, of a Web page. In this way, all search requests with similar (and usually fuzzy) structure characteristics can be grouped into a WIST, which is implemented as an intelligent software agent using FL, NN, GA, and other technologies to automatically find all relevant Web pages based on the relevance inferred from structure rules and user-submitted keywords, and rank them in a desired way. Essentially, a WIST agent uses a TSK-based Fuzzy Neural Network (FNN) to infer the desirability. The agent is intelligent because it can 1 learn to get better parameters of FNN, 2 learn to get better structure of FNN, and 3 learn to define structure characteristics by adding/modifying structure rules.

Hybrid Fuzzy Neural Networks for Protein Secondary Structure Prediction

Yuchun Tang, Preeti Singh, Yanqing Zhang, Chun-Dar Lu, Irene Weber

This paper presents a new fuzzy neural network method for predicting protein secondary structures. Three protein secondary structures (i.e., alpha helix, beta sheet and coil) are used for protein classification. The traditional orthogonal encoding scheme on the other hand takes an inordinate amount of convergence time to train neural networks. To solve this problem, the new encoding scheme for the various amino acids is proposed to map 20 amino acid symbols to numerical values in [0, 1] based on relevance degrees among the 20 amino acids. For comparison, the conventional multi-layer neural network with the Back Propagation (BP) learning is also used to do protein secondary structure prediction. The training data sets and testing data sets are generated differently from public protein data in order to verify prediction performance objectively.

The hybrid neural network system consists of (1) three independent neural networks for predicting alpha helix, beta sheet and coil respectively, and (2) the final decision-making system that classify the current protein sequence into one class with the maximum prediction output of the three prediction outputs generated by the three independent neural networks.

The hybrid conventional neural network has three independent neural networks using BP learning. The accuracy for predicting alpha helix and beta sheet can reach 70% and 60%, respectively for a training data set of 200 and more. The prediction accuracy for coil can reach 77.7% for a training data set of 70 sequences. The new hybrid fuzzy neural network has three independent fuzzy neural networks using the new fuzzy learning. The initial simulation results have been generated by the hybrid fuzzy neural network. The accuracy for predicting alpha helix and beta sheet can reach 80.0% and 70.0%, respectively for a training data set of 200 and more. The prediction accuracy for coil can reach 88.8% for a training data set of 70 sequences.

In summary, the new hybrid fuzzy neural network is more effective than the hybrid conventional neural network in terms of prediction accuracy and learning speed. In the future, seven protein secondary structures will be used in prediction. The hybrid fuzzy neural network will be improved by adding new intelligent techniques such as genetic algorithms and kernel-based learning.

World Knowledge and Fuzzy Logic

Lotif A. Zadeh

Informally, world knowledge is the knowledge acquired through experience, education and communication.  World knowledge has a position of centrality in human cognition, and especially in summarization, assessment of relevance, deduction and search.

Centrality of world knowledge in human cognition entails its centrality in issues related to web intelligence. In the existing literature, world knowledge is dealt with through the use of knowledge-representation methods based on bivalent logic. The problem is that much of world knowledge, e.g., "California has a temperate climate" and " Most Finns are honest " is perception-based and/or dispositional. Such knowledge does not fit the conceptual structure of bivalent logical systems -- logical systems which are intolerant of imprecision, uncertainty and partial truth.

A thesis which is advanced in this talk is that to deal with world knowledge what is needed is fuzzy logic and, more particularly, computational theory of perceptions, dispositional logic and precisiated natural language. An outline of fuzzy-logic-based approach to world knowledge is presented and illustrated by examples.

From Search Engines to Question-Answering Systems

Lotif A. Zadeh

Search engines, with Google at the top, have many remarkable capabilities. But what is not among them is the deduction capability - the capability to synthesize an answer to a query by drawing on bodies of information which are resident in various parts of the knowledge base. It is this capability that differentiates a question-answering system, Q/A system for short, from a search engine. Upgrading a search engine to a Q/A system is a complex, effort-intensive, open-ended problem. Semantic Web and related systems for upgrading quality of search may be viewed as steps in this direction. But what may be argued, as is done in the following, is that existing tools, based as they are on bivalent logic and probability theory, have intrinsic limitations. The principal obstacle is the nature of world knowledge.

The centrality of world knowledge in human cognition, and especially in reasoning and decision-making, has long been recognized in AI. The Cyc system of Douglas Lenat is a repository of world knowledge. The problem is that much of world knowledge consists of perceptions. More specifically, perceptions are f-granular in the sense that (a) the boundaries of perceived classes are fuzzy; and (b) the perceived values of attributes are granular, with a granule being a clump of values drawn together by indistinguishability, similarity, proximity or functionality. What is not widely recognized is that f-granularity of perceptions put them well beyond the reach of computational bivalent-logic-based theories.

Dealing with world knowledge needs new tools. A new tool which is suggested for this purpose is the fuzzy-logic-based method of computing with words and perceptions (CWP), with the understanding that perceptions are described in a natural language. A concept which plays a key role in CWP is that of Precisiated Natural Language (PNL). It is this language that is the centerpiece of our approach to reasoning and decision-making with world knowledge.

A concept which plays a key role in organization of world knowledge is that of an epistemic (knowledge-directed) lexicon (EL). Basically, an epistemic lexicon is a network of nodes and weighted links, with node i representing an object in the world knowledge database, and a weighted link from node i to node j representing the strength of association between i and j. The name of an object is a word or a composite word, e.g., car, passenger car or Ph.D. degree. An object is described by a relation or relations whose fields are attributes of the object. The values of an attribute may be granulated and associated with granulated probability and possibility distributions. For example, the values of a granular attribute may be labeled small, medium and large, and their probabilities may be described as low, high and low, respectively. Relations which are associated with an object serve as PNL-based descriptions of the world knowledge about the object. For example, a relation associated with an object labeled Ph.D. degree may contain attributes labeled Eligibility, Length.of.study, Granting.institution, etc. The knowledge associated with an object may be context-dependent. What should be stressed is that the concept of an epistemic lexicon is intended to be employed in representation of world knowledge - which is largely perception-based - rather than Web knowledge, which is not.

In conclusion, the main thrust of the fuzzy-logic-based approach to question-answering which is outlined in this abstract, is that to achieve significant question-answering capability it is necessary to develop methods of dealing with the reality that much of world knowledge is perception-based. Dealing with perception-based information is more complex and more effort-intensive than dealing with measurement-based information. In this instance, as in many others, complexity is the price that has to be paid to achieve superior performance.

BISC Decision Support System (BISC-DSS)

Masoud Nikravesh

A key component of any autonomous system is a decision module, which is capable of handling large volume of data at high speed and high reliability. This paper will be focused, in the main, on the development of a decision module, which is capable of functioning in an environment of imprecision, uncertainty and imperfect reliability.  There will be two principal tasks: Task A, which will be aimed at the development of novel methods of analysis and design: and Task B, which will be focused on the development of a decision-support system for ranking of decision alternatives. Following is an outline of the principal issues, which will be addressed in Task A.

In this study, we introduced the BISC decision support system as an alternative for ranking and predicting the risk which utilizes an imprecise and subjective process. The BISC decision support system key features are 1) intelligent tools to assist decision-makers in assessing the consequences of decision made in an environment of imprecision, uncertainty, and partial truth and providing a systematic risk analysis, 2) intelligent tools to be used to assist decision-makers answer “What if Questions”, examine numerous alternatives very quickly and find the value of the inputs to achieve a desired level of output, and 3) intelligent tools to be used with human interaction and feedback to achieve a capability to learn and adapt through time.


Funding for this research was provided by the British Telecommunication (BT) and the BISC Program of UC Berkeley.


M. Nikravesh (2001) Credit Scoring for Billions of Financing Decisions, Joint 9th IFSA World Congress and 20th NAFIPS International Conference. IFSA/NAFIPS 2001 "Fuzziness and Soft Computing in the New Millenium", July 25-28, 2001

M. Nikravesh and B. Azvine, “Fuzzy queries, search, and decision support system”, Soft Computing, vol. 6, pp. 373–399, 2002.

M. Nikravesh, B. Azvine, R. Yager and Lotfi A. Zadeh (2003b) "New Directions in Enhancing the power of the Internet", Editors:, to be published in the Series Studies in Fuzziness and Soft Computing, Physica-Verlag, Springer (August 2003)

M. Nikraves and Lotfi A. Zadeh, "Fuzzy Logic an the Internet", to be published in the Series Studies in Fuzziness and Soft Computing, Physica-Verlag, Springer (August 2003)

M. Nikravesh, B. Azvine, R. Yagar and Lotfi A. Zadeh, "New Directions in Enhancing the power of the Internet", to be published in the Series Studies in Fuzziness and Soft Computing, Physica-Verlag, Springer (Dec 2003)

Zadeh L. A. (1999), From Computing with Numbers to Computing with Words -- From Manipulation of Measurements to Manipulation of Perceptions, IEEE Transactions on Circuits and Systems, 45, 105-119, 1999.

Zadeh, L. and Kacprzyk, J. (eds.) (1999a), Computing With Words in Information/Intelligent Systems 1: Foundations, Physica-Verlag (1999a).

Zadeh, L. and Kacprzyk, J. (eds.) (1999b), Computing With Words in Information/Intelligent Systems 2: Applications, Physica-Verlag,(1999b).

Zadeh, L.A. (1996) Fuzzy Logic = Computing with Words,” IEEE Trans. on Fuzzy Systems (1996) 4, 103-111.

Zadeh, L. A. and M. Nikravesh (2002), Perception-Based Intelligent Decision Systems, AINS; ONR Summer 2002 Program Review, 30 July-1 August, UCLA.

BISC-DSS: Evolutionary Computing-based Multi-Aggregator Fuzzy Decision Trees and optimization

Souad Souafi-Bensafi, Masoud Nikravesh

In the world of information processing, we are faced with increasingly complex multi-domain problems containing either real-world or computer-generated data. To consider these problems the classical data processing tools may not be sufficient and more advanced approaches need to be developed. A unified approach based on Soft Computing will help in solving such problems by combining methodologies (Fuzzy Logic, Neuro-Computing, Evolutionary Computing and Probabilistic Reasoning), which collectively provide a foundation for the conception of Intelligent Systems.

Our aim is to develop intelligent computing techniques that address the problem of multi-criteria decision making dealing with subjective and imprecise data. This kind of problem requires conception of intelligent systems able to replace a human with expertise in a specific domain in a decision making process. So, the intelligent system should take into account the subjective and imprecise character of data on the one hand, and represent the user or expert’s preferences and knowledge on the other hand. For this purpose, we developed a generic multi-criteria model based on fuzzy logic concepts for decision support systems. Our goal is to build such a model by 1) fitting real-world data and 2) representing the preferences of specific-domain users or experts. Toward this end, we used Evolutionary Computation techniques. Initially, we worked on a first order aggregation model and performed its learning using Genetic Algorithms, in which these preferences have been represented by a weighting vector associated with the variables involved in the aggregation process. This has been used in a specific application related to university admissions. Then, we developed a more advanced multi-aggregation model based on a hierarchical decision trees and for the learning process of this model, we developed a technique inspired from Genetic Programming. In this model tree nodes represent aggregators, terminals or leaves correspond to variables, and weight values are added to the children branches for each aggregator. The aggregation result overall the variables is then obtained by running recursively the root aggregator of the tree.

The parameters characterizing this multi-aggregation model are aggregators, weights and their combination in form of a tree structure. In this case, the learning process has to find the optimal combination of these parameters based on training data. In this learning process, the evolution principle remains the same as in a conventional GP but the DNA encoding needs to be defined according to the considered problem. We need to define a more complex tree structure representing the multi-aggregation model. In addition to the weights that have to be added to the classical tree, the nodes that represent the aggregators require a variable number or arguments. This is because the number of arguments cannot be known before the tree creation. Therefore, during the evolution process, trees are generated randomly by selecting aggregators for the nodes, and at the same time, the corresponding numbers of arguments are randomly chosen. Moreover, weight values are fixed randomly in each branch of the tree during its creation.

These encoding properties allow a large search space to solve our problem. If we need to simplify this tree structure according to some application constraints, we can add these constraints to the problem specification and they will be checked during the tree generation.

We pursue this work by considering many other applications and we aim at carrying out the multi-aggregation model base on decision trees. We defined this model as a basis for a more general and a more complex form of the considered problems which operates with linguistic variables. Our approach is a first attempt toward the use of the Computing with Words and Perceptions.

Web-Based BISC Decision Support System

Gamil Serag-Eldin, Masoud Nikravesh

Most of the existing search systems ‘software’ are modeled using crisp logic and queries. We introduce fuzzy querying and ranking as a flexible tool allowing approximation where the selected objects do not need to match exactly the decision criteria resembling natural human behavior. The model consists of five major modules: the Fuzzy Search Engine, the Application Templates, the User Interface, the Database and the Evolutionary Computing. The system is designed in a generic form to accommodate more diverse applications and to be delivered as stand-alone software to academia and businesses.

In this study, we introduced fuzzy query, fuzzy aggregation, evolutionary computing and the BISC decision support system as an alternative for ranking and predicting the risk for credit scoring, university admissions, and several other applications, which currently utilize an imprecise and subjective process. The BISC decision support system key features are 1) intelligent tools to assist decision-makers in assessing the consequences of decision made in an environment of imprecision, uncertainty, and partial truth and providing a systematic risk analysis, 2)intelligent tools to be used to assist decision-makers answer “what if”, questions examine numerous alternatives very quickly and find the value of the inputs to achieve a desired level of output, and 3) intelligent tools to be used with human interaction and feedback to achieve a capability to learn and adapt through time In addition, the following important points have been found in this study: 1) no single ranking function works well for all contexts, 2) most similarity measures work about the same regardless of the model, 3) there is little overlap between successful ranking functions, and 4) the same model can be used for other applications such as the design of a more intelligent search engine which includes the user's preferences and profile (Nikravesh, 2001a and 2001b).


Funding for this research was provided by the British Telecommunication (BT) and the BISC Program of UC Berkeley.


Fagin R. (1998) Fuzzy Queries in Multimedia Database Systems, Proc. ACM Symposium on Principles of Database Systems, pp. 1-10.

Fagin R. (1999) Combining fuzzy information from multiple systems. J. Computer and System Sciences 58, pp 83-99.

J. R. Koza, Genetic Programming : On the Programming of Computers by Means of Natural Selection, Cambridge : MIT Press 1992, 819 pages.

John H. Holland. Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. MIT Press, 1992. First Published by University of Michigan Press 1975.

Mizumoto M. (1989) Pictorial Representations of Fuzzy Connectives, Part I: Cases of T-norms, T-conorms and Averaging Operators, Fuzzy Sets and Systems 31, pp. 217-242.

Nikravesh M. (2001a) Perception-based information processing and re-trieval: application to user profiling, 2001 research summary, EECS, ERL, University of California, Berkeley, BT-BISC Project. (http://zadeh.cs.berkeley.edu/ & http://www.cs.berkeley.edu/~nikrav

es/ & http://www-bisc.cs.berkeley.edu/).

Nikravesh M. (2001b) Credit Scoring for Billions of Financing Deci-sions, Joint 9th IFSA World Congress and 20th NAFIPS International Conference. IFSA/NAFIPS 2001 "Fuzziness and Soft Computing in the New Millenium", Vancouver, Canada, July 25-28, 2001.

Masoud Nikravesh and Ben Azvine (2002), Fuzzy Queries, Search, and Decision Support System, Journal of Soft Computing, Volum 6, # 5, August 2002.

Masoud Nikravesh, B. Azvine, R. Yagar, and Lotfi A. Zadeh (2003) "New Directions in Enhancing the power of the Internet", Editors:, to be published in the Series Studies in Fuzziness and Soft Computing, Physica-Verlag, Springer (August 2003).

BeMySearch: Concept-Based Information Retrieval and Search Engine

Tomohiro Takagi, Masoud Nikravesh

Since a fuzzy set is defined by enumerating its elements and the degree of membership of each element, we can use it to express word ambiguity by enumerating all possible meanings of a word, then estimating the degrees of compatibilities between the word and the meanings.

Based on this approach, we have proposed using conceptual fuzzy sets (CFSs) to represent the various meanings of a concept that change dynamically depending on the context. A CFS (is realized as neural networks in which a node represents a concept and a link represents the strength of the relation between two (connected) concepts. The activation values agreeing with the grades of membership are determined through this associative memory. In a CFS, the meaning of a concept is represented by the distribution of the activation values of the other nodes. The distribution evolves from the activation of the node representing the concept of interest.

This presentation will start with my motivation to propose CFSs and algorithm to generate CFSs. It will describe how it works to represent the context dependent meaning of a word and to measure a conceptual distance between documents. Next, information filtering and image search (Google-Based Search Engine for Multimedia Data) will be introduced as its applications to information retrieval using capability of conceptual matching. Finally we will introduce our approach to enhancing CFSs based on brain architecture.

BeMySearch: Conceptual Search Engine and Navigation for Google™ and Yahoo!, Web Intelligence

Masoud Nikravesh, Tomohiro Takagi

In this presentation, first we will present the role of the fuzzy logic in the Internet. Then we will present an intelligent model that can mine the Internet to conceptually match and rank homepages based on predefined linguistic formulations and rules defined by experts or based on a set of known homepages. The FCM model will be used for intelligent information and knowledge retrieval through conceptual matching of both text and images (here defined as "Concept"). The FCM can also be used for constructing fuzzy ontology or terms related to the context of the query and search to resolve the ambiguity. This model can be used to calculate conceptually the degree of match to the object or query. We will also present the integration of our technology into commercial search engines such as Google ™ and Yahoo! as a framework that can be used to integrate our model into any other commercial search engines, or development of the next generation of search engines.

Constructing Fuzzy Thesauri for the WWW: application to BeMySearch

Martine De Cock, Sergio Guadarrama, Masoud Nikravesh

The remarkable growth of the World Wide Web (WWW) since its origin in the 1990's calls for efficient and effective tools for information retrieval. Attempting to deal with the overwhelming amount of information provided on billions of webpages nowadays does not necessarily imply that we have to develop entirely new technologies from scratch. In the 1970's and 1980's initial research was performed on the retrieval of information from modest text collections, using fuzzy relations to represent dependencies between terms on one hand, and between terms and documents on the other. Since then the fuzzy mathematical machinery (i.e. fuzzy logical operators, fuzzy similarity measures, operations with fuzzy relations) has come of age. We revisit the “old" information retrieval strategies and brush them up with new insights to prepare them for a great challenge: search in the biggest collection of text known by mankind.

Automated Consumer Profiling using Fuzzy Query and Social Network Techniques

F. Olcay Cirit, Sema Alptekin, Masoud Nikravesh

Decades of research in the field of Social Network Analysis have established the existence of a phenomenon known as homophily: the tendency of people of similar behavior and social condition to co-associate. One can observe this phenomenon on a small scale when girls in the same sorority tend to favor the same brands of products, while on a national scale one finds that people in high-income brackets overwhelmingly tend to have friends who are also in high income-brackets.

Many internet companies today, among them America Online™, Friendster™, and Yahoo!™, maintain large databases of information about their users, and, if they do not already, are at least in a position to capture information about how each user connects socially to the other users of the system. A web site could gather this information directly by correlating lists of friends that each user provides, or indirectly by correlating usage of online discussion forums and chat rooms. This project examines the possibility of applying the social connectivity information thus gleaned to profile potential consumers of a product or service. Also discussed are techniques for using the BISC/DSS, a fuzzy query-based decision support system, to implement an automated profiling.

For ease of conceptualization, we compare being a consumer of a product or service to being a member of a social club. The task of consumer profiling, then, can be thought of as the search for potential new club members. Alternatively, one can ask the following question: given a set of connected persons, which of them are most likely to belong to club X? One possible approach to solving these problems involves constructing a profile of the members already in the club, and then calculating the similarity of each user to the profile. Static information about the individuals in the club, such as age, income, interests, etc., could be supplemented with social network metrics such as connectedness, betweeness, closeness, etc. to provide a social grounding for the profile. Such an approach is distinctly suited for use with the DSS because the definition of the profile corresponds exactly with a fuzzy query.

Targeted internet advertising and customer relationship management could make profitable use of the consumer profiling techniques outlined here. A proposed implementation of an automated consumer profiling system that integrates with the BISC/DSS is examined in detail, including a discussion of methods for collecting social connectivity information, and an assessment of the pitfalls of profiling.

On Similarity Measures and Pruning Strategies for Fuzzy Ontologies

Andrey Kolobov, Daniel Kuo, Martine De Cock, Masoud Nikravesh

We critically examine the applicability and usefulness of mathematical models for the construction of fuzzy term-term relations based on fuzzy similarity and fuzzy inclusion measures. Much attention is paid to realistic pruning algorithms of generated fuzzy relations, both semantically as well as regarding execution time.

Performing region detection using simple fuzzy agents

Yonathan Asfaw, Martine De Cock, Masoud Nikravesh

Fuzzy Reasoning can be applied in work that requires decision-making. The use of simple agents is also gaining ground in computer science. We joint the two concepts to employ simple agents that have minimal interaction with one another to attempt to carry out region detection in images using fuzzy reasoning.