23.2.1994
Artificial Intelligence Laboratory 
==================================

Research in Artificial Intelligence in Ljubljana began in 1972
at the Computer Science Department of the Jozef Stefan Institute
(IJS). The laboratory was formally founded in 1979 (initially as
AI group). The laboratory is lead by Professor Ivan Bratko, a well
known researcher in the AI field with over 100 international
publications. His book "Prolog Programming for Artificial 
Intelligence" (Addison-Wesley) was translated into major world
languages (English, German, Italian, French, Slovenian, Japanese,
Russian).

In the first 10 years, the emphasis was on theoretical research
which provided a solid background for later application projects.
In 1982, the development and implementation of AI tools started
and soon resulted in practical applications. By now, over 60
projects were completed in different domains, including business
and management (e.g. evaluation of investments, projects and
enterprises), technical domains (e.g. metallurgy, geology, civil
engineering) and medicine. AI laboratory offers commercial
versions of empirical and decision making systems as well as
expertise - meaning over 150 man-years in this area. 

The majority of applications are based on Assistant, an
inductive learning system, and DEX, a computer aided decision
making system. Several other systems are research prototypes 
indicating practical applicatibility, among them GINESYS, LINUS, 
m-FOIL, MARCUS, RETIS. These systems have been compared to the
best available on the market and have shown performance at or
above their level.

Members of the laboratory actively cooperate in several 
faculty activities in Slovenia and abroad.

In 1993, the Artificial intelligence laboratory participated in the
following projects funded by the Slovenian Ministry of Science
and Technology: Automated Knowledge Synthesis, Knowledge
Engineering and Expert Systems - Knowledge-Based Systems.

The AI laboratory also participated in the following international
research projects: ESPRIT III Basic Research Project no. 6020
Inductive Logic Programming, Esprit III Network of Excellence in
Computational Logic, and Esprit III European Network of Excellence in
Machine Learning. The AI laboratory is the coordinator of the 
PECO92 project ILPNET - Inductive Logic Programming Pan-European
Scientific Network. The Third International Inductive Logic
Programming Workshop was organized at Bled in April 1993.

The scope of the Artificial Intelligence Laboratory is in the 
following topics:
(a) Inductive logic programming
(b) Probabilistic learning in noisy domains
(c) Evolutionary algorithms 
(d) Integration: multistrategy learning and multiple knowledge
(e) Knowledge synthesis for modelling, control, diagnosis and design
(f) Knowledge synthesis and decision support

Detailed description:

(a) Inductive logic programming

Inductive Logic Programming (ILP) is a research area in the
intersection of machine learning and logic programming. It aims
at a formal framework and practical algorithms for inductive
learning in first-order logic. An important aspect of the 
practical algorithms is the handling of noisy data.

The integrated learning system LINUS, that can learn 
attribute-value and relational descriptions, was used to learn 
medical diagnostic rules from real-life noisy data. 
The basic LINUS algorithm was extended with the possibility to 
introduce new variables, thereby extending the class of ILP 
problems that can be solved by transforming to attribute-value 
form, where existing mechanisms can be used to handle noise.
The extended algorithm was used to prove the PAC-learnability 
of ij-determinate and constrained nonrecursive logic programs.

A study of noise-handling mechanisms was also conducted in the
context of multiple predicate learning. The study also showed
that existing ILP systems, such as FOIL and GOLEM, are
inappropriate for the problem of learning multiple predicates.
A new algorithm, MPL, was developed, which solves many of the 
problems present in the existing systems mentioned.

The ILP system MARKUS, which brings several improvements over MIS,
was used to solve problems from software engineering, innovative 
design and interpreting results of discrete-event simulation. 

(b) Probabilistic learning in noisy domains

Developed methods for handling probablility measures and noise,
implemented in systems RETIS, were applied in ecologic modelling,
drug design and in particle physics.

In eceologic modelling  regression trees, constructed by RETIS,
were used to predict quantity and growth of algae in the lagoon
of Venice. Since genereated knowledge was nontrivial and
consistent with expert knowledge we judge that the first results
are encouraging.

In drug design the biological activity of a drug was predicted
from its structural activities.

In particle physics, indentification of subatomic paticles on
the basis of data obtained from electromagnetic calorimeter was
carried out. During the analysis it was foud out that most of
the 18 attributes used initially can be omitted without
seriously affecting the classification accuracy. USing less
attributes yields smaller and more understandable decision
trees, which are preferred by the experts.

(c) Evolutionary algorithms

In 1993, we have continued the investigation of evolutionary
algorithms as an optimisation method for solving problems in the
fields of dynamic systems control and attribute machine learning.
The work resulted in a PhD thesis, in which genetic algorithms
have been applied in solving these problems, and several
international publications.

In the research of methods for dynamic systems control, an
interactive genetic algorithm has been developed to tune the
parameters of a PID controller to regulate a demanding nonlinear
and unstable process on a testing laboratory device. The algorithm
performs parameter optimisation through a series of trials on a
real-world device. The approach makes it possible to involve the
optimality criteria preferred by human operators and improves the
results obtained with classical methods from the control theory.

A genetic-algorithm-based machine learning system, called Gallus,
has also been developed. The system employs the cover algorithm
and, given a set of learning examples described by attributes and
classes, generates a decision list to classify new examples.
Binary encoding of the classification rules is used, and
traditional genetic operators crossover and mutation are applied
in the evolution of the rules. Experiments on real-world domains
have confirmed that a genetic algorithm can well serve as a basis
of a machine learning system.

(d) Integration: multistrategy learning and multiple knowledge

Research in multistrategy learning and multiple knowledge 
representations has in 1993 moved towards the Japanese project 
Real-World Computing, and memory-based approach, one of 
the central areas at IJCAI93 in France.

In our laboratory, in-depth modeling of multiple knowledge has 
been continued and intensified. Formal average-case analyses of 
multistrategy models strongly indicate that important 
improvements can be obtained in this way. Formal analyses have 
been derived especially for two classifiers. With growing number 
of classifiers, the classification accuracy tends to grow up to a
point, and then starts to decrease depending on parameters.

At Acroni, Iron and Steel Works, Jesenice, a universal 
integrating schema has been designed and implemented. For the 
moment it encapsulates four systems, however, is fully open for
integrating other data-compatible systems. Expert systems at 
Acroni Jesenice have been in regular use for years. The new 
system has achieved a couple of percent better accuracy in tests
and is now in regular use for one year.

Functionally similar environment for integrating systems has
been designed at the Institute. It integrates around 10 systems
for empirical learning, inductive logic programming and
statistics. Together, there are around 100.000 lines of original
program code.

A prototype of an Intelligent Operating Interface (IOI) for 
VAX/VMS has been designed. It is based on the memory-based 
approach, and implements some of the properties of the 
future-generation interfaces. For example, it adapts and learns 
from the VAX/VMS-user communication. It has shown surprisingly
big advantages of the new approach - by extensive use of huge and
quick computer memories it is possible to implement many 
properties of intelligent behaviour. Instead from programs, 
intelligence stems from huge amounts of data.

(e) Knowledge synthesis for modelling, control, diagnosis and design

We apply knowledge synthesis techniques to the problems of modelling,
control, diagnosis and design of systems, including dynamic systems.
Besides being interested in their own right, these sophisticated areas
serve for practical verification and evaluation of the developed
techniques. 

Following the idea of using alternative semantics in ILP, 
the systems QMN and LAGRANGE were developed, which discover
qualitative and quantitative laws governing dynamical systems from 
example behaviours. QMN and LAGRANGE exhaustively search the space
of equations over the system variables and their derivatives, 
testing the validity of equations by using linear regression. 
They successfully reconstructed the models of several dynamical 
systems from simulated data. The GOLDHORN system was also developed, 
which adds to LAGRANGE the capability of handling noisy 
measurements. GOLDHORN reconstructed the models of two real 
dynamical systems from measurements of their behaviour. 

The problem of planning with qualitative models of dynamic 
systems was defined. It was shown that a QSIM-type model of a dynamic
system can be defined in the event calculus formalism, as a
domain of abductive planner APEC.

A three-phase framework for learning dynamic system control was
developed. First, a genetic algorithm is employed to derive
control rules encoded as decision tables, without using prior
knowledge about the controlled system. Next, the learned rules
are automatically transformed into comprehensible form by means
of inductive learning. Finally, a genetic algorithm is applied
again to optimise the numerical parameters of the induced rules.
This framework suggests the decomposition of control synthesis
problem, in which learning without prior knowledge as the most
time-consuming step remains just for the worst case. Otherwise,
control strategy as an intermediate objective can be obtained
either by taking into account prior knowledge at the symbol
level, or by extracting operator's implicit strategy from the
recordings of his/her performance. In our work, both cases have
been investigated. Recently, the efforts have concentrated in
the domain of container crane control. For this purpose, a
numerical simulator as well as a small-scaled physical model
have been developed.

(f) Knowledge synthesis and decision support

In the area of artificial intelligence methods for supporting
decision-making processes, the research was focused to the integration
of qualitative multi-attribute decision models with the quantitative
ones, which are based on numeric attributes and analytically
represented utility functions. We developed a new method that
automatically transforms quantitative models to their quantitative
counterparts. The method retains all the advantages of the qualitative
models, such as transparency, comprehensibility, effective knowledge
acquisition and explanation, but, on the other hand, considerably
improves the sensitivity of models. This is specially important in
decision problems that involve a large number of options.

The new method was successfully applied in several complex real-world
decision-making problems. In the collaboration with the Ministry of
Science and Technology of the Republic of Slovenia we developed an
expert system for the evaluation of research and development projects.
In the years 1992 and 1993, about 1500 projects were evaluated by this
system. The second problem was the allocation of housing loans, for
which a similar system was developed and used to evaluate about 6000
applications for loans. We also developed some other prototype decision
support systems for strategic planning, portfolio analysis and
classification of waste disposals.
============================================