The aim of the paper is to present our ongoing project on modeling
virtual enterprise.
We propose a simple model of "a networked world"
represented
by a computer network where workstations play roles of computing
factories with specific inputs and outputs.
Each computing factory is managed by one agent called {\em factory agent.}
There is global demand for some types of final products
(computations)
so that in order to manufacture final products the factories must
be formed in one
or several competing supply chains that in turn are organized into
enterprises.
The work of enterprise formation and integration is supposed to be
performed by mobile agents created by the factory agents.

keywords:
agent-based manufacturing, multi-agent systems,
enterprise formation

Modern visualization techniques like tour plots, augmented biplots,
parallel coordinate plots -- combined with some other, auxiliary evaluations and
using an appropriate GUI -- permit to recognize the shape of the multivariate data
cloud containing points representing subsequent rows from the data table X, in
particular to see outliers, if any.
We demonstrate the usefulness of these techniques considering the structure of
earnings for some data extracted from the Eurostat New Cronos data base.

Keywords: Multivariate outliers, Grand tour, Biplot, Parallel coordinate plot

The paper presents a new method and the corresponding program
of presentation of bayesian belief networks.
The belief network can be viewed and updated via World Wide Web.
Consistency checks are possible. Edge removal and insertion operations
are done in an `intelligent way' that is corrections of valuations are
carried out automatically in a user-friendly way.
The corresponding program is implemented as a Java applet at the front
end,
and is backed by some Java applications at the server site.

Keywords
Knowledge representation, knowledge acquisition from the user, belief
networks.

Paper discusses changes introduced in engineering by the
rapid development of communications media. In particular,
potential advantages of distributed knowledge-based
tools for design are highlighted. Such tools enable the
designer to assemble an object using intelligent components
that posses the ability to satisfy constraints and to adjust
themselves to neighbours.
Depositories of intelligent components accessible through
Internet will replace traditional catalogues used to

Keywords:
Distributed data-bases, concurrent design,
intelligent agents.

The paper describes hybrid pattern recognition system composed of computer
generated hologram (CGH) in optical part and artificial neural network
(ANN) in digital part. The process of CGH regions optimization with the use
of rough sets theory (RST) elements and evolutionary stochastic programming
is presented. As an example of such hybrid system the optic fiber
distortion classifier is described.

Keywords:
Ring-Wedge Detectors, Computer Generated Holograms, Rough Sets,
Evolutionary Optimization, Neural Networks

The determination of an IIR filter structure is a classic approximation problem. This paper contains a description of the application of genetic algorithm to task of design filter with arbitrary magnitude and group delay. The new method makes it possible obtained filter with lower order as compared to traditional method. The lowpass, highpass, bandpass and bandstop filters can be independently designed.

Keywords: Signal processing, filter design, application of genetic algorithm.

This paper presents the results of a research on methods of analytical data sets clustering and graphic presentation. In the research, the SAHN algorithm (SAHN- Sequential, Agglomerative, Hierarchical, and Non-overlapping method) has been applied as a clustering tool. The visualization procedures are used not only to present the results of data classification but also to estimate the chosen clustering methods. The SAHN methods were applicated in a computer system (VVT module in the SCANKEE system) and are used in unsupervised learning, classifying new objects and predic-ting their properties.

Key words: data classification, data clustering, intelligent infor-mation systems, applications of system science in chemistry

Methodology of extraction, optimization and application of sets of logical rules
is described. The three steps are largely independent. Neural networks are used for initial
rule extraction, local or global minimization procedures for optimization, and Gaussian
uncertainties of measurements are assumed during application of logical rules. A tradeoff
between rejection/error level is discussed. This methodology has been applied to a number
of benchmark and real life problems with very good results.

Keywords: Neural networks, logical rules, fuzzy rules, optimization, medical diagnosis.

Several methods for feature selection and weighting have been implemented
and tested within the similarity-based framework of classification methods. Features are ex-
cluded and ranked according to their contribution to the classification accuracy in the cross-
validation tests. Weighting factors used to compute distances are optimized using global
minimization procedures or search-based methods. Our experiments show that, for some
datasets, these methods give much better results than classical nearest neighbor methods.

Keywords: similarity based methods, kNN, optimization, feature selection, classification

The theories of dynamical systems and partial differential equations offer
various models for realistic problems that often exhibit spatial and temporal
changes. However, one of the significant difficulties lies in both the resolution
and implementation of such equations. Indeed, most natural systems are discrete
in nature and their evolution generates very complex behaviours. A discrete
mathematical idealization, based on cellular automata, provides an alternative
approach to modelling such processes including spatial expansion phenomena.
The aim of this paper is to summarize some mathematical notions recently
introduced in order to describe various phenomena closely linked to population ecology.
Such models have many advantages, particularly in their successful implementation
to a wide variety of ecological applications.

In this paper we investigate the problems of data retrieval in full text databases utilizing flexion-based languages, such as
Polish language. Such databases prove to more difficult in processing than those created with isolating languages (i.e. English language)
due to inflexion and resulting inconsistent sentence structure. Data retrieval in full text databases is also more complex than in highly
organized databases as for satisfying results some kind of semantic analysis must be employed in most applications. We describe an
experimental data retrieval system, which addresses these problems using combination of inflexion database,
simple algorithmic stem finder and thesaurus-based semantic analyzer. This system uses UNESCO developed ISIS full text database
system, but the modifications necessary for use in World Wide Web environment are also discussed.

Keywords: natural language processing, full-text databases, thesauri, flexion-based languages, web searching

Computer technology with its powerful tools for calculation and analysis,
for information storage and retrieval and for modeling and simulation - has
become increasingly important, as an aid in the design process. However
these tools demonstrate only a small fraction of the potential possibilities
of using computers in design. A designer still wastes time and his attention
on certain useless tasks, as far as the design process is concerned, i.e.
linking data between separate and specialized programs and trying to keep
the track of the design progress. The use of control architectures could
help designers to deliver higher quality solutions faster and more
efficiently. The most important aspects of control architectures will be
described in the context of the design process of mobile crane breaking
system.

Keywords:
Design process, computer-aided design, blackboard architecture, agents.

The problem of discovering association rules has received research
attention. Several fast algorithms have been developed. The key problem
in discovery of association rules is discovering frequent itemsets.
Typical algorithms for solving this problem operate in a breadth-first
search strategy using a bottom-up and/or top-down direction. This paper
presents a new efficient algorithm for discovery of association rules in
large databases. The main idea of the method is based on heuristic
depth-first search strategy.

KEYWORDS:
data mining, association rules, depth-first search, heuristic.

We studied two prenatal data sets and two other medical data
sets. Our objective was to increase sensitivity (accuracy of preterm
birth prediction) by changing the rule strength for the preterm birth
class. Two criteria for choosing the optimal rule strength are
discussed: the greatest difference between the true-positive and false-
positive probabilities and the maximum profit.

KEYWORDS: Preterm birth, system LERS, data mining, rule induction,
machine learning, rough set theory, classification of cases.

Progress of own research on the algorithm aimed at hierarchical reasoning over uncertain and/or not complete knowledge base in rule-based systems is briefly dealt with.

Keywords: case-based reasoning, distributed processing

In this paper a model is constructed for games with communication under incomplete information. The model, based on natural language processing techniques, is developed and tested on a case of Bridge card play bidding. The approach proposed here emphasizes the information processing nature of such games. Finally, some conclusions from implementation of a
prototype bidding system and a new methodology for performance evaluation of such systems are presented.

keywords:
knowledge based systems, natural language processing, game solving, games with incomplete information, conversational agent, reasoning, performance evaluation, Horn clauses.

The general problem of off-line adaptation of the Binary Ge-
netic Algorithm (BGA) is introduced. An example of such an adaptation:
a class of Correlational Adaptation Methods (CAMs) is proposed. The
main idea of a CAM is that it uses a mapping called a measurement
function as an assessment of the BGA's efficiency. An example of such a
measurement function is described and two examples of CAMs: a mod-
ified ``trials and errors'' method and a modified genetic meta-algorithm
(metaBGA) are shown. Finally, experimental results with the metaBGA
for four kinds of test fitness functions, where we use a code permutation
as the adapted parameter, are presented.

Keywords. genetic algorithm, adaptation, optimization, code permutation.

We present an approach to a certain diagnostic problem. Our goal was
to build a knowledge database that contains diagnostic rules making possible
recognition of shafting form of the turbine set. We consider problem of selec-
tion diagnostic signals and its attributes, and also definition of threshold values of
this attributes.

Keywords: machinery diagnostic, machine learning, rotating machinery.

Key words: aggregation; clustering; decomposition; dependence; grade
correspondence analysis; Kendall's $tau;$ regularity of dependence; Spearman's $rho.$

The paper presents designing a timetable for the city tram service using a genetic algorithm (GA). Definition of the task in the terms of GA, i.e., representation of individuals, the coding schema and the evaluation function is described. A case studies of two real tram services - in Wroc³aw and Poznań are presented.

Key words: genetic algorithm, optimisation, city tram service

A new method of learning decision rules from databases, which uses
an evolutionary algorithm, is proposed. The main difference
between our approach and the others described in the literature is
the way of processing of continuous-valued attributes. Most
decision rule learners process separately these attributes when
searching for threshold values, which may decrease the
performance. In contrast to them, our method globally searches for
threshold values for all continuous-valued attributes in the rule
at the same time. The initial results of the experiments on some
real-life datasets are presented.


Keywords:
machine learning, evolutionary algorithms, decision rules

This paper describes AQ-PM, a system for partial memory learning,
which determines and memorizes representative concept examples, and
then uses them with new training examples to induce new concept
descriptions. Our approach uses ``extreme'' examples that lie at
the boundaries of current concept descriptions. We evaluated
the system by applying it to synthetic and real-world learning
problems. In the experiments, the partial memory learner notably
reduced memory requirements for storing examples at the slight
expense of predictive accuracy. The system also performed well
when tracking concept drift.

Key words: concept learning, rule induction, partial instance memory models

This paper presents next step of our work to find different connections between rough sets and evidence
theory. We would like to insist on way of finding subsets of attributes which are good direction for discernment of
decision's classes. This problem makes sense for the inconsistence decision table.

Keywords: rough sets, evidence theory

We develope a data-mining technology to deal with two discrete, multiple
choice variables. Our aim is to assess correlation between these variables.
Within the model of the Rasch type defined earlier (Matuszewski, 1998) an
interpretation is introduced which enable possibilities for further
theoretical research, improvement of numerical algorithms and practical
application.


Keywords:
maximum likelihood estimation, conditional estimation, exponential family,
multiple dichotomy, multiple choice question, data mining

In this paper an application of decision rules to function representation in
reinforcement learning is described. Rules are generated incrementally by method based on
rough set theory from instances recorded in state-action-Q-value memory. Simulation
experiment investigating the performance of the system and results achieved are reported.

Keywords: Rough Set Theory, Decision Rules, Reinforcement Learning, Q-learning

The diagnostics of machinery is nowadays aided by expert systems
which require knowledge of the machine to be diagnosed. The paper deals
with a new method of knowledge acquisition from examples, especially
useful in technical diagnostics of machinery where complex structure of
a set of technical states often occurs. The approach depends on optimal
decomposition of the set of examples. The selection criterion may
concern the highest performance of the family of state classifiers
determined by the contents of the knowledge base acquired by means of
induction. An example concerning an application of the method for
acquisition of diagnostic knowledge on rotating machinery is shown, too.

Keywords:
machinery diagnostics, complex technical states, knowledge
acquisition, learning from examples.

This work presents an approach to identification of dynamic systems
which exploits a network of dynamic neurons. Each neuron consists of an adder
module, a linear dynamic system --- Infinite Impulse Response (IIR) filter, and
nonlinear activation module. Thus its activation depends on actual inputs as well
as inputs and outputs at the previous time-moment. Up-to-date training algo-
rithms, based on classical back propagation, suffer from entrapment in local min-
ima of a time-varying error function. An evolutionary search with soft selection
and forced direction of mutation is proposed as a training algorithm to overcome
these difficulties.

Keywords: neural modeling, dynamic neural model, evolutionary algorithms

The paper deals with the object-oriented techniques used to design and
develop expert system shells. The object-oriented approach is used on
several levels: level of modules, modules' architecture and a knowledge
base. Finally the expert system shell Consus is described, as a realization
of this approach.

Keywords: expert systems, object-oriented technology.

This paper presents an application of reinforcement learning (RL) to
the school timetable problem. In this approach a learning agent
performs heuristic search of a space of timetables in order to find
an acceptable one and improves the heuristic being used on-line.
After each timetable modification the agent is either punished for
prolonging the search process or it receives a reward if a solution
is found. Since it tries to maximize the reinforcement values
received, it learns to solve the problem as quickly as possible.
Because only \emph{evaluative} information is needed, applying RL
algorithms requires from the designer significantly less knowledge
about the problem domain than in standard methods where a teacher is
involved in the process of learning.
In our paper we give a short
overview of reinforcement learning and focus on formulating a school
timetable problem in terms of RL which includes defining such
elements of a problem as a set of states, a set of operators, and a
reinforcement function. Several computational experiments
investigating the performance of the system and some key features of
this approach are reported.

Keywords: Machine Learning, Reinforcement Learning, Heuristics,
Problem Solving, Timetable Problem

This paper describes a machine learning system being developed in our team,
using different inference strategies to create the description of a concept.
The control of the learning process complies with the assumptions presented by
R. S. Michalski in his work Inferential Theory of Learning [1]. Horn clauses have
been employed as the knowledge representation language. The system described
in this paper is a starting point for further research on a system using the
confidence factors (CF) and operating on noisy data. The structure and concept of
the system are similar to formerly developed multistrategy systems, especially to
MTL-JT and DISCIPLE.

keywords: Machine Learning, Multistrategy Systems, Deduction, Analogy, Induction,
Abduction.

Verification of knowledge bases has emerged as a significant problem in the development of expert systems (ES). Although the basic
verification concepts are shared by software engineering and knowledge engineering, verification methods of conventional software are
not directly applicable to ES and the new, specific methods of verification are required.
We argue that knowledge bases verification can not be delayed until the final knowledge base realization and we suppose that
verification should be performed incrementally and should be included into the development process. In this paper we introduce the
main assumptions of dynamic verification approach. Next we briefly describe the current state of work on the an assistant tool for
incremental knowledge base building and verification.

Keywords:
expert systems, knowledge bases, verification, validation.

The paper presents an interactive rule discovery system based on a
new algorithm DataExplore. The user can express requirements to
the strength, length and level of confidence of the rule, as well
as to the syntax of condition part. In succeeding iterations he
focuses his attention first on more general then on more specific
rules. Analysis technical diagnostics of mechanical vehicles by
means of implementation of DataExplore is discussed.

keywords:
knowledge discovery, decision rules, interactive systems,

The paper addresses the problem of analysing information tables which
contain objects described by both attributes and criteria, i.e. attributes
with preference-ordered scales. The objects contained in those tables,
representing exemplary decisions made by a decision maker or domain expert,
are often classified into one of several classes that are
preference-ordered. Analysis of such data using the classic rough set
methodology may produce unsatisfactory results, as the original rough set
approach is not able to discover inconsistencies coming from consideration
of criteria, e.g. product quality, market share or debt ratio. The paper
presents the framework for analysis of both attributes and criteria and a
fast algorithm for generating reducts in this framework. The algorithm is
comprehensively evaluated in an experiment with real-life data sets.

Keywords:
intelligent information systems, rough sets theory, multi-attribute and
multi-criteria classification, reducts of attributes and criteria.

Key words: computer- intensive methods, contingency table, graphical display, make-believe outlier, monotone dependence, occupational mobility, scatterplot

This paper gives a formal definition of Collective Intelligence (C-I)
and defines its IQ measure (IQS). This has been allowed by the
application of a specific molecular, quasi-chaotic model of computations
- the Random PROLOG Processor (RPP). This approach is expected to work
for a spectrum of social structures of beings: bacterial and insect
colonies, social animals up to human social structures. Using this
theory some social phenomena can be explained as optimization toward
higher IQS. The definition of C-I is based on the assumption that it is
a temporary and variable property of a social structure, initialized
when individuals organize or acquire the ability to solve more complex
problems than best individuals can. This property is amplified when the
social structure improves its synergy. The definition covers both when
C-I results in physical synergy or in logical cooperative
problem-solving.

Keywords:
Collective Intelligence, IQ, social structure, computational model,
information molecules, computational space, PROLOG, problem-solving,
Brownian movements.

Most real-world applications operate in dynamic environments. In
such environments often it is necessary to modify the current solution due to
various changes in the environment (e.g., machine breakdowns, sickness of em-
ployees, etc). Thus it is important to investigate properties of adaptive algorithms
which do not require re-start every time a change is recorded.
In this paper non-stationary problems (i.e., problems, which change in time) are
discussed. We describe different types of changes in the environment. A new model
for non-stationary problems and a classification of these problems by the type of
changes is proposed. A brief review of existing applied measures of obtained results
is also presented.

This paper presents an evolutionary approach to infer natural language analyzer from legal and illegal examples of natural language text in the form of isolated sentences. It provides theoretical bases for the use of two classes of evolutionary computation, that is evolutionary programming and genetic programming, that support automated inference of fuzzy automaton-driven analyzer of natural language. This analyzer, called fPDAMS (fuzzy nondeterministic pushdown automaton with associative memory access), works in the stratificational knowledge representation system. In the evolutionary programming (EP) each chromosome is a representation of a transition graph of fPDAMS. The key mechanism used in this approach is asexual mutation. Genetic programing (GP) can be applied to natural language analyzer if a mapping is established between the point-labeled tree used in GP and the transition graph of the automaton. The method of mapping is based on the encoding technique called cellular encoding and proposed by F.Gruau.

Keywords:
natural language processing, fuzzy automata, evolutionary computation

The paper presents the problem of using neural network for military vehicle classification on the basis of ground vibration. One of the main element of the system is a unit called geophone. This unit allows to measure amplitude of ground vibration in each direction for certain period of time. The value of amplitude is used to fix the characteristic frequencies of each vehicle. If we want to fix the main frequency it is necessary to use Fourier transform. In this case the fast Fourier transform FFT was used. Because the neural network (Radial Basis Function network) was used, the learning set has to be prepared.
Please find attached the results of using RBF neural network such as: example of learning, validation and test sets, structure of the networks and learning algorithm, learning and testing results.

Keywords: neural networks, classification, signal processing

The work consists of two parts. In the first part the idea of genetic
programming is presented and the basic elements of a genetic programming system
are described. In the second part, considering a selected example, we describe
the results of investigations of the influence of program grammars on the
efficiency of genetic programming.

Keywords:
Genetic programming, genetic algorithms, grammars

The paper reveals a few differences between the rough sets and statistics, and discovers the principle of accurate database mining, which is inherent with the rough sets. The rough sets database mining system precisely computes optional non-linear n-side relationships existing within N variables. As such, the rough sets is an excellent tool to find combinations of significant variables and their intervals playing a role in a database. The mathematical model of the database mining based on the rough sets is shown and the results present numbers which may not be revealed by using standard statistical procedures. Statistics mines directly to 2-side relationships, and these relationships differ from those existing in reality in a database. The problem of detecting optional n-side relationships within N variables is NP-complex. Insufficient computational tools at the time of the development of statistics can account for the situation. The new rough sets database mining system presented in this article solves the NP-completeness of mining to all patterns (significant combinations of intervals within N variables), without splitting the simultaneous and unknown n-side relationships into pairs. This new system has been developed by Smart Machines.

Key Words: ARS, accurate database mining, rough sets, survival analysis, inaccuracy of statistics, n-side relations, categorical analysis, patterns, non-statistical inference.

Robot-discoverers and other intelligent systems should inter-
act with the physical world in complex, yet purposeful and accurate ways.
Knowledge representation which is internal to a computer lacks empirical
meaning and thus it is insufficient for the investigation of the external
world. We argue that operational definitions are necessary to provide
empirical meaning of concepts, but they have been largely ignored by
the research on automation of discovery. In this paper we reconstruct
the scientific mechanism by which operational definitions are created
and we make several steps towards the implementation. Individual oper-
ational definitions can be viewed as algorithms that operate in the real
world. They can and they should be improved in the course of interac-
tion with the real world, so that their accuracy is improved. We explain
why many operational definitions are needed for each concept and how
different operational definitions of the same concept can be empirically
and theoretically equivalent. We argue that all operational definitions of
the same concept must form a coherent set and we define the meaning of
coherence. No set of operational definitions is complete. We argue that
expanding the operational definitions is one of the key tasks in science.
Among many possible expansions only a very special few lead to a sat-
isfactory growth of scientific knowledge. While our examples come from
natural sciences, where the use of operational definitions is especially
clear, operational definitions are needed for all empirical concepts. We
briefly argue their role in a robot-discoverer and in database applications.