Learning Innovations

THE MODEL OF ASIS FOR PROCESS CONTROL APPLICATIONS

Plamena Andreeva, Tatiana Atanasova, Jordan Zaprianov

plamena@iusi.bas.bg, tatiana@iusi.bas.bg, zaprian@bgcict.acad.bg
http://www.cl.acad.bg/directory/tech/icsr.html
Institute of Control and System Researches, BG

ABSTRACT

This paper describes a new model of information system, called Adaptive Self-learning Information System (ASIS). Designing an intelligent system needs decision-making process in a fuzzy environment. We build an information system, which stores the different alternatives about a process control. A dynamic method for reasoning in the knowledge base is discussed. Fuzzy rules determine the central component of the information processing. Fuzzy cluster approach based on Bezdek (FCM) is used to classify the input data and to receive the rules. The construction of ASIS is proposed to make the process of decision making easy, at low cost and automated. The proposed model of ASIS is suitable for industrial process control as for training systems.

1. INTRODUCTION

The state of the art in networked scientific computing has progressed very slowly. For solving a given problem it is necessary to identify the best resources from specialised scientific software servers, libraries, repositories distributed across the Internet. To establish high-level problem solving inter-networked environments it is need to have a distributed knowledge base and intelligent brokers for retrieval required information. The aim of the development of our Adaptive Selflearning Information System (ASIS) is to support the decision making process in knowledge based system. ASIS is intended to work in the automatic control system design environment with engineering knowledge-based management as an integral part.

The problem is that each of the existing systems has its own user interface and individual language to model the knowledge. In this paper, the proposed ASIS performs the estimation of the "best" alternative about the process control, and suggests the most appropriate decision for a given criterion. It has the following structure: knowledge data base (KDB) part with stored different alternatives about the process control; decision-making part; and control part as output of the "best" solution. Data processing is made due to fuzzy presentation of knowledge.

The idea of fuzzy control is to formulate the control algorithm by logical rules. Fuzzy rules determine the central component of the information processing. There are three major approaches to the derivation of the rules of a process control:

derivation of rules from the experience-based knowledge of the process;
using linguistic description (fuzzy model of the process under control) to derive the rules;
on the base of a conventional process model (nonlinear) making a fuzzy version of Sliding Mode Control.

We use the second one approach and apply some of the concepts used in control theory to develop and formalize methods for constructing fuzzy logic controller (FLC). By control application of the linguistic approach (Hirota, 1994) and (Driankov, 1993) fuzzy logic is used to convert heuristic control rules, as stated by a human operator, into an automatic control strategy. In this sense the FLC is nothing more than a linguistic model of the human operator strategy and is a decision model. In contrast to fuzzy set theory, probability theory control and information theory accept that imprecision can be equated with randomness. Coupling fuzzy logic with expert system technology provides a mechanism for producing fuzzy models that address important classes of problems in decision support systems.

2. THE MODEL OF ASIS

In the presented model of ASIS we deal with heuristic (knowledge from expert) by describing the rules linguistically. ASIS performs the estimation of the "best" alternative for a given criterion, and suggests the best decision. The information system is adaptive, this means the given input value is followed and the difference is reduces to fit the criterion. It is self-learning because even by new unknown input the fuzzy classifier scheme proceeds and executes the rule to find a new fact. The basic structure of ASIS is given on figure 1.

Figure 1. Basic structure of ASIS

In the Data Base block the a-priory input control information is stored. Then the rule firing due to Rule Base is proceeded and the proposed solution is offered. The fuzzy reasoning on the Knowledge Base (KB) is applied iterative, so new facts to KDB are added, which allows stepwise refinement in ASIS.

2.1 Structure and Elements of ASIS

Designing an intelligent system needs decision-making process in a fuzzy environment. We want to build an information system, which stores the different alternatives about a process control. Then we can estimate the "best " alternative for a given criterion. The way in which the input information is processed in ASIS is following the 3 blocks diagram, shown on figure 2. This is a query view of Inference mechanism from the ASIS basic structure. The input flow of information passes through the next phases:

linguistic variable base;
fuzzy classifier and
decision making block.

Figure 2. Information processing in ASIS.

The linguistic variable descriptor is a first stage, which produces a linguistic data representation. The classification is made with Min-max composition. It is used to derive a new fact from the KDB according to the applicable rules. The third stage Decision-making produces the output of the system. It gives the estimation of "best" matches for a given criterion. The content of the KDB is changed cyclically during the data processing. The facts may have time tags, so that the time of their insertion into the KDB can be determined.

The proposed mechanism for describing fuzzy variables is a modification of the classification problem. The most important feature of this approach is that it makes possible to combine implicit and explicit priorities. It is dynamic method, the reasoning to the knowledge base is applied iterative and new facts to the KDB are added.

2.2 Information Processing in Knowledge Base

When we have only vague information about the values of the attribute a_i, the matching between the data and the criteria leads to a fuzzy number. The simplest form for a fuzzy database is the attachment of a membership value (numeric or linguistic) to each tuple T_j in KDB. In the relation the membership value denotes the strength of the dependency between the key and the attribute. An essential aspect of most fuzzy relational DB is that domain values d_j1, d_j2, ... d_jm need not be atomic. An interpretation x [a₁, a₂,... a_n] of tuple Tj = [d_j1, d_j2, ... d_jm] is any value assignment such that a_i d_ij for all _i. The space of interpretations is the set cross product D1 x D2 x ... D_n. In an ordinary relational DB a tuple is equivalent to its interpretation.

2.3 Fuzzy Rules in ASIS

Fuzzy rules determine the central component of the information processing. Fuzzy cluster approach is used to classify the input data and to receive the rules. The original idea of using fuzzy sets in clustering techniques was introduced by (Bezdek, 1981) and consists in the fuzzification of the partition functional.

The similarity measure between k-th point X_k and i-th collection of typical points (named centroid V) is then defined as the euclidean distance D_k,i²

he whole development is to be recast in a fuzzy setting, where the data are to be classified into C clusters. The points X_k are weighted with their fuzzy membership. The clusters V_i (centroids) are defined as the weighted fuzzy sum of the X_k points:

The standard variables are restricted to shapes which allow and overlap of no more than two terms. This leads to a variable definition in which each term is determined by exactly one parameter, the base value of the term's maximum. Then the result of the clustering is new combined data with shortest distance. To qualify the existing knowledge information the method of fuzzification fuzzy clustering mean (FCM) is used as shown on figure 3.

Figure 3. Fuzzy clustering based on FCM.

The first part fuzzification converts current value of the process state variable into fuzzy set to be compatible with the presentation of the process state variable in the rule base. The rule base is a set of production rules. The next part Inference engine performs rule firing. The compositional rule of inference uses a fuzzy relation to represent explicitly the connection between two fuzzy propositions. It can be considered as a special case of General Modus Ponens (GMP). By individual based inference - the overall value of the control output is to be compute based on individual contributions of each rule. The last part Defuzzification converts the set of modified control output values into a single point-wise value.

3. DYNAMIC METHOD

The main advantage of an on-line information system, working in network environment involving intranet and/or Internet is the possibility of dynamic changes and updates. To design computer-based decision supporting information, such as ASIS, a dynamic method for retrieval and adding of new rules needs to be used. We deal with fuzzy dynamic algorithm for clustering and estimation of existing knowledge base. Then the information, stored in ASIS is processed cyclically until given criterion is fitted.

The fuzzy query processing has an imprecise predicate. In conventional relational DB operations are performed only on boolean predicates. We try to use this and make an extension of SQL language. First, from initial fuzzy query we search a boolean query with given x level, where the received tuples are calculated to be with membership degree >=x. Here x is called lambda cut and represent the degree of matching . The membership value of a tuple represents the best matching interpretation.

3.1 Fuzzy Relational Data Base Operation

Data processing is made due to fuzzy presentation of knowledge. ASIS executes all possible rules in four steps:

find the most specific rule to be fired;
execute the rule on the last data facts (data are stored in DB with time tags)
do this (in cycle) while no rules are left, or until criterion is satisfied;
every rule can be fired only once

In this paper we consider the discrete decision space (situations in which there are countable many actions from which the "optimal" has to be chosen). Utilities (cost function) are modeled as fuzzy sets, in the framework of linguistic variables. The method for linguistic description of fuzzy variable (Andreeva, 1995) is used to process the information in ASIS.

4. KNOWLEDGE PRESENTATION

The knowledge representation method used in the design of an expert system has direct impact on its development, efficiency, speed and maintenance. There are many approaches to represent Knowledge. The main four of them are: 1) first-order logic, 2) production system, 3) semantic nets and 4) frame. Here we are applying the approximate reasoning based on Production system approach.

The information knowledge is codified in rules, which consist of linguistic variables. A priori possibility distributions are defined (over some base variable) for each linguistic value assumed by the elementary expressions, and therefore it is possible to assign to each one of them an 'evaluation'. Then we assign to each formula the judgement, made by the observer, about the truth of some information , on the bases of gathered knowledge.

4.1 Operation on Knowledge Base

The fuzziness in KDB can be considered on all components i) fuzzy description of entities; ii) fuzzy relationship between entities and iii) fuzzy attributes presentations. The operations performed on knowledge base are:

Store facts - check that it does not conflict with previously stored facts;
Retrieve facts - performs queries;
Inference, i.e. the retrieval of implicit knowledge;
Return definition of concepts;
Remove facts from the knowledge base.

4.2 Method for Fuzzy Variable Description

A fuzzy set A has membership function [0,1] which takes crisp values. Now we have to translate this linguistic variable in fuzzy number. The operation on this fuzzy number can be unary, called modifier, and binary (or tree-ary), called composition of the two (or three) fuzzy number. A fuzzy relation R is a mapping over the Cartesian product between fuzzy sets A and B.

R: A x B -> [0,1]

For each element x from A and y from B, x_R(x,y) express the strength of their bond; the higher the more closer to 1. One of the first attempts at the quantification of word meaning was made by Mousier. He hypothesized that the meaning of a word may be considered as containing two components: a constant, reflecting the overall meaning value and a variable component representing the variation in the meaning of the word due to context.

Fuzzy relational models express dependencies between system's variable in terms of fuzzy relations rather than functions. In terms of KBD system, relations are relevant for representing the formal association between antecedents (if-part) and consequent (then-part). Fuzzy Relations in different product spaces can be combined with each other by the operation "composition". Different versions of "composition" have been suggested which differ in their results. The max-min composition has become the best known. However the max-product or max-average composition leads to results that are more appealing.

CONCLUSION

In this paper a new model of ASIS is proposed, where we seek to make the process of decision-making easy, at low cost and automated. ASIS performs the estimation of the "best" alternative, and suggests the best decision. Data processing is made due to fuzzy presentation of knowledge. The information system is adaptive, this means the given input value is followed and the difference is reduces to fit the criterion. It is selflearning because even by new unknown input the fuzzy classifier scheme proceeds and executes the rule to find a new fact. A dynamic method for reasoning in the knowledge base is applied iterative and new facts to the KDB are added, which allows stepwise refinement in ASIS. The fuzziness in Relational DB can be considered on all components a) fuzzy description of entities; b) fuzzy relationship between entities and c) fuzzy attributes presentations.

Current work is aimed at refinement and extension of ASIS application in control theory. As a further related work it will be interesting to apply Object Oriented Data Base (OODB) with fuzzy queries.

REFERENCES

Andreeva Pl., (1995). Inexact information system and its application to approximated reasoning, Proc. of BIOMATH'95, August 23-27, Sofia.

Bezdek J.C., (1981). Pattern recognition with fuzzy objective function algorithms, Plenum Press, New York.

Cox E., (1999). Fuzzy Systems Handbook, Second Edition, Academic Press, San Diego, CA.

Driankov D., Hellendoorn H., Reinfrank M., (1993). An Introduction to fuzzy control, Springer Verlag, Berlin Heidelberg.

Hirota K., (Ed), (1994). Industrial Applications of Fuzzy Technologies, Springer Verlag, Tokyo, Berlin.

Kickert Walter J.M. (1978). Fuzzy theories on decision-making, Frontiers in systems research, Vol.3, Kluwer Boston Inc., Netherlands.

Li H., Gupta M., (1995). Fuzzy logic and intelligent systems, Kluwer Acad. Press.

Miyamoto Sadaaki, (1990). Fuzzy sets in information retrieval and cluster analysis, Kluwer Academic Publishers, Netherlands.

Palm R., Driankov D., Hellendoorn H., (1997). Model based Fuzzy Control, Springer-Verlag, Berlin, Heidelberg.

Pedrycz W., (1996). Fuzzy modeling. Paradigms and Practice, Kluwer Academic Publishers.

Quinlan M., in Shapiro S., Eckroth D., (1987). Encyclopedia of Artificial Intelligence, John Wiley & Sons Inc., New York.

Zimmermann Hans-J., (1987). Fuzzy sets, decision making, and expert systems, Kluwer Academic Publishers, Boston.