Seminarji Slovenskega društva za umetno inteligenco (1999)

26.1.1999
Angus R. Simpson (University of Adelaide, Dept. of Civil and Environmental Engineering):
Optimisation of Water Distribution Systems Using Genetic Algorithms
2.3.1999
Marko Grobelnik (IJS):
Predstavitev članka: Identifying Hierarchical Structure in Sequences: A Linear-Time Algorithm
9.3.1999
Nada Lavrač (IJS):
Experiments with Noise Filtering in the Diagnosis of Coronary Artery Disease
16.3.1999
Dragica Smrke, Ivan Bratko, Janez Demšar, Vlado Stankovski :
Nova metodologija za medicinske raziskave?
30.3.1999
Dunja Mladenić (IJS):
Avtomatska kategorizacija dokumentov z uporabo strojnega učenja
8.4.1999
Kalyanmoy Deb (Kanpur Genetic Algorithms Laboratory (KanGAL), Indian Institute of Technology Kanpur, India):
Multi-Objective Optimization Using Genetic Algorithms
13.4.1999
Igor Grabec (Fakulteta za strojništvo, Univerza v Ljubljani):
Kompleksnost eksperimentalnih podatkov in optimalno modeliranje fizikalnih zakonov
22.4.1999
Hendrik Blockeel (Department of Computer Science, Katholieke Universiteit Leuven, Belgium):
First-Order Logical Decision Trees and their Application to Water Quality Prediction
1.7.1999
Hiroshi Motoda (Division of Intelligent Systems Science, The Institute of Scientific and Industrial Research, Osaka University, Japan):
Discovery of First Principle Equations from Numeric Data
1.7.1999
Nobuhiro Inuzuka (Department of Intelligence and Computer Science, Nagoya Institute of Technology, Japan):
An Extension of Inductive Logic Programming Based on Fuzzy Logic Programming
31.8.1999
Edward R. Vrscay (Department of Applied Mathematics, University of Waterloo, Canada):
Fractal Image Compression to ODEs: Solving Inverse Problems Using Contraction Maps
14.9.1999
Bogdan Filipič (IJS in Univerza v Ljubljani, Fakulteta za strojništvo):
Nov mejnik v razvoju evolucijskega računanja
21.9.1999
Katsuhiko Nakamura (Tokyo Denki University, Saitama, Japan):
Graph-Theoretic Analyses of Board Phases in the Game of Go
23.9.1999
Igor Kononenko (FRI), Tatjana Zrimec (FRI) in sodelavci:
Dosedanji rezultati raziskav s Kirlianovo kamero
28.9.1999
Sven Erik Jorgensen (Royal Danish School of Pharmacy, Copenhagen, Denmark):
Application of Thermodynamics to Describe the Development of Ecosystems
16.11.1999
Aleks Jakulin (FRI):
Vizualizacija
23.11.1999
Vladislav Rajkovič (FOV in IJS):
Nekateri izzivi inteligentnih sistemov za pomoč pri odločanju
14.12.1999
Ljupčo Todorovski, Sašo Džeroski (IJS):
Meta Level Classification Trees


Angus R. Simpson (University of Adelaide, Dept. of Civil and Environmental Engineering):
Optimisation of Water Distribution Systems Using Genetic Algorithms

Normal design practice for design of water distribution systems is to use computer hydraulic simulation models. A trial and error approach is used. In this approach, locations and sizes of components (for example - pipes, pumps, tanks and pressure reducing valves) in the water distribution system are selected by the designer. The computer simulation model is run for various demand load cases (peak hour and fire); the actual pressures are analysed; and then changes to the network are made to attempt to minimise cost while satisfying all the design criteria. Optimisation approaches have been developed over the last 30 years but virtually none of these approaches have made it into usage by consulting and practicing engineers. A relatively new optimisation technique is called genetic algorithm optimisation. Research work at the University of Adelaide over the last 9 years has developed a comprehensive approach to the optimisation of the design and operation of water distribution systems. This seminar will describe how the genetic algorithm technique is applied to water distribution system design. Some case studies where substantial savings in capital costs using the genetic algorithm technique will be described.


Marko Grobelnik (IJS):
Predstavitev članka

C.G. Nevill-Manning, I.H.Witten: Identifying Hierarchical Structure in Sequences: A Linear-Time Algorithm. Journal of Artificial Intelligence Research 7, 67-82, 1997.


Nada Lavrač (IJS):
Experiments with Noise Filtering in the Diagnosis of Coronary Artery Disease

This paper (written by Gamberger, Lavrač, and Grošelj) presents a series of noise detection experiments in a medical problem of coronary artery disease diagnosis. The following algorithms for noise detection and elimination are tested: a saturation filter, a classification filter, a combined classification-saturation filter, and a consensus saturation filter. The distinguishing feature of the novel consensus saturation filter is its high reliability which is due to the multiple detection of potentially noisy examples. Reliable detection of noisy examples is important for the analysis of patient records in medical databases, as well as for the induction of rules from filtered data, representing genuine characteristics of the diagnostic domain. Medical evaluation in the problem of coronary artery disease diagnosis shows that the detected noisy examples are indeed noisy or non-typical class representatives.


Dragica Smrke, Ivan Bratko, Janez Demšar, Vlado Stankovski :
Nova metodologija za medicinske raziskave?

Predstavitev in diskusija (brainstorming) o tezi:
Strojno učenje omogoča znanstveno metodologijo, ki je bistveno manj restriktivna od tradicionalno privzete metodologije (na primer na področju medicinskih raziskav), ki predpostavlja uporabo statistike.


Dunja Mladenić (IJS):
Avtomatska kategorizacija dokumentov z uporabo strojnega učenja

Na seminarju bom predstavila kategorizacijo dokumentov z uporabo metod za učenje na tekstovnih podatkih, ki so bile razvite v moji doktorski disertaciji. Pogledali bomo uporabo Yahoo hierarhije za avtomatsko kategorizacijo spletnih dokumentov. Pomembno je, da predlagane metode upoštevajo posebnosti problemske domene. Ena od posebnosti je neuravnovešena distribucija po vrednostih razreda. Druga je nesimetrična cena napačne klasifikacije, ki je razvidna le implicitno iz narave problema. V našem primeru bi lahko rekli, da je bolje napačno klasificirati negativen primer, kot napačno klasificirati pozitiven primer. Te posebnosti upoštevamo pri izbiranju podmnožice atributov in ocenjevanju kvalitete modela. Razen tega imamo opravek z velikim številom atributov in primerov, kar zahteva učinkovito izvedbo predlaganih metod.


Kalyanmoy Deb (Kanpur Genetic Algorithms Laboratory (KanGAL), Indian Institute of Technology Kanpur, India):
Multi-Objective Optimization Using Genetic Algorithms

Many real-world search and optimization problems are naturally posed as non-linear programming problems having multiple objectives. Due to the lack of suitable solution techniques, such problems are artificially converted into a single-objective problem and solved. The difficulty arises because such problems give rise to a set of Pareto-optimal solutions, instead of a single optimum solution. It then becomes important to find not just one Pareto-optimal solution but as many of them as possible. Classical methods are not efficient because they require repetitive applications to find multiple Pareto-optimal solutions and in some occasions repetitive applications do not guarantee finding distinct Pareto-optimal solutions. The population approach of genetic algorithms (GAs) allows an efficient way to find multiple Pareto-optimal solutions simultaneously in a single run.

In this talk, one implementation of a GA which uses non-domination concept will be discussed. Simulation results on a number of test problems and on a couple of engineering design problems will show the efficacy of the method. Clearly, GAs (or other evolutionary search methods) have a niche in solving multi-objective optimization problems compared to classical methods. This is why multi-objective optimization using GAs is getting a growing attention in the recent past. In this talk, a number of future challenges in the research of multi-objective optimization will also be discussed.


Igor Grabec (Fakulteta za strojništvo, Univerza v Ljubljani):
Kompleksnost eksperimentalnih podatkov in optimalno modeliranje fizikalnih zakonov

Na seminarju bo opisano statistično modeliranje fizikalnih zakonov na osnovi izmerjenih podatkov. Model tvori Parzenova cenilka gostote porazdelitve verjetnosti, v kateri je jedro določeno s sipalno funkcijo instrumenta. Z entropijo informacije porazdelitve verjetnosti sta nato definirani kompleksnost in redundanca eksperimentalnih podatkov. Kriterijska funkcija modela je sestavljena iz redundance in napake predikcije, z njenim minimumom pa je opredeljen optimalni model opazovanega pojava. Za iskanje optimuma je uporabljeno kreiranje in anihiliranje parametrov modela. Numerični primeri kažejo, da je formiranje optimalnega modela podobno procesu kondenzacije.


Hendrik Blockeel (Department of Computer Science, Katholieke Universiteit Leuven, Belgium):
First-Order Logical Decision Trees and their Application to Water Quality Prediction

The main ideas underlying the decision tree learner Tilde are explained. These are:

  1. a general approach to predictive induction that we call "predictive clustering" and that generalises over classification, regression and certain types of clustering;
  2. the notion of first-order logical decision trees, which are the first-order equivalent of classical decision trees.
The Tilde system is thus able to induce from structural data first order theories that allow both numerical and symbolic predictions, but can also be used for, e.g., unsupervised learning, pattern completion, and simultaneous prediction of multiple variables.

As an illustration of the various uses Tilde can be put to, we discuss some of its applications in the domain of river water quality prediction.


Hiroshi Motoda (Division of Intelligent Systems Science, The Institute of Scientific and Industrial Research, Osaka University, Japan):
Discovery of First Principle Equations from Numeric Data

We show that there is a method that ensures to derive the first principle from numeric data. Notion of scales and an interesting property that is deduced by dimensional analysis are the basis of the method. These two together with simple mathematics can constrain the form of admissible relations. The method works for phenomena for which nothing is known about the number of equations that are needed to describe them and no knowledge of dimensions of the variables involved is available. Many of the known first principles that involves several tens of variables and a few tens of equations have been derived by numerical experiments with noisy data.


Nobuhiro Inuzuka (Department of Intelligence and Computer Science, Nagoya Institute of Technology, Japan):
An Extension of Inductive Logic Programming Based on Fuzzy Logic Programming

This talk presents an extension of framework of inductive logic programming (ILP) based on fuzzy logic programming (FLP). FLP treats intermediate truth between true and false, and has possibility to describe more sophisticated knowledge than the normal LP. I am going to discuss a framework of induction of fuzzy logic programs as inductive FLP (IFLP). I also present a top-down algorithm to solve IFLP problems, the algorithm which is given be extending FOIL algorithm. I also discuss that the IFLP framework can treat attribute-value form knowledge more effectively than the normal ILP framework.


Edward R. Vrscay (Department of Applied Mathematics, University of Waterloo, Canada):
Fractal Image Compression to ODEs: Solving Inverse Problems Using Contraction Maps

Fractal-based image coding methods seek to approximate a target image v(x) with the fixed point attractor function u(x) of a contractive fractal transform operator T that acts upon a uitable metric space (F,dF) of image functions. The action of the fractal transform T on a function u involves (1) creating a number of contracted copies of u, (2) modifying these copies and (3) recombining them to form a new function Tu. The problem is to find the "best" such operator T. The parameters that define T - the so-called fractal code - are then used to represent the target image v. The approximation u is generated by the iteration sequence un+1 = T un, where u0 in F is a suitable "seed" (for example, a blank screen). Under suitable conditions, the computer storage of the fractal code requires much less memory than that of the original image. The result, therefore, is lossy image compression since full accuracy is generally not achieved.

More generally, fractal-based coding involves the approximation of elements of an appropriate metric space by fixed points of contractive operators on that space. As expected, Banach's Fixed Point Theorem is the basis of this method along with an important corollary that has been called the "Collage Theorem."

It has been our goal to look for other, possibly "nonfractal," inverse problems in which these methods can be employed. One such example is a simple class of inverse problems involving ordinary differential equations: Given a function x(t) (which may be an interpolation of a set of experimental data points (xi,ti )), find an ODE x. = f(x,t) that best admits x(t) as an exact or approximate solution, where f is restricted to a class of functional forms, e.g. polynomial. (For example, a population biologist may wish to find the best Lotka-Volterra system that admits a set of experimental results as solution.) In this case, we use the Picard contraction mapping and, as in fractal-based methods, find the Picard map T (in turn defined by f) that maps the target x(t) as close as possible to itself.

This talk, aimed at a general audience, will outline the major points behind the "approximation by fixed points problem," using fractal coding as an illustrative example. We then move on to the above-mentioned inverse problem for ODEs and show how the Picard integral operator is used to construct optimal vector fields. Some simple examples in one and two dimensions are then presented.


Bogdan Filipič (IJS in Univerza v Ljubljani, Fakulteta za strojništvo):
Nov mejnik v razvoju evolucijskega računanja

Začetki evolucijskega računanja segajo v šestdeseta leta, ko so prvi raziskovalci uporabili koncepte biološke evolucije kot hevristike za računalniško reševanje problemov. Ločeno so bile razvite različne metode: genetski algoritmi, evolucijske strategije in evolucijsko programiranje. To delo pa ob svojem nastanku ni naletelo na ustrezen odmev in je za skoraj dve desetletji utonilo v pozabo. Šele v osemdestih letih je s splošnim razvojem računalništva in uveljavitvijo subsimbolnih metod doživelo svoj preporod. Pomembni mejniki v nadaljnjem razvoju evolucijskega računanja so bili prva mednarodna konferenca o genetskih algoritmih leta 1985, izid Goldbergove knjige leta 1989, začetek sodelovanja med različnimi šolami evolucijskega računanja leta 1991 in začetek izhajanja revije Evolutionary Computation (MIT Press) leta 1993. V zadnjem desetletju področje dosega eksponentno rast tako po številu objav in znanstvenih srečanj kot projektov in praktičnih rezultatov. Letos poleti sta bila v ZDA dva velika mednarodna kongresa, posvečena evolucijskim algoritmom: Congress on Evolutionary Computation (CEC-99) in Genetic and Evolutionary Computation Conference (GECCO-99). Zdi se, da zaradi svojega obsega, prikazanih rezultatov in načrtov za prihodnost predstavljata nov mejnik v uveljavljanju področja. Na seminarju bomo predstavili vsebinske poudarke in novosti z obeh srečanj. Dodajamo tudi nekaj koristnih povezav:

Konference:
CEC-99: http://garage.cps.msu.edu/cec99/
GECCO-99: http://www-illigal.ge.uiuc.edu/gecco/
CEC-2000: http://www.natural-selection.com/eps/cec2000/
GECCO-2000: http://www.genetic-algorithm.org/GECCO2000/gecco2000mainpage.htm

Evolucijsko računanje:
FAQ: http://www.cs.purdue.edu/coast/archive/clife/FAQ/www/
ENCORE: ftp://ftp.cs.wayne.edu/pub/EC/Welcome.html
GA Archive: http://www.aic.nrl.navy.mil/galist/
EvoNet: http://www.dcs.napier.ac.uk/evonet/Coordinator/evonet.html
Literatura: http://www.genetic-programming.org/books.html


Katsuhiko Nakamura (Tokyo Denki University, Saitama, Japan):
Graph-Theoretic Analyses of Board Phases in the Game of Go

Determining life or death for groups of stones by analyzing board phases is most important in computer Go. We have been developing two methods of analyzing life and death of the groups. The first method is based on Euler's Formula for connected planar graphs, by which the number of regions is calculated by the numbers of vertices and edges. The other method is to use labeled graphs called semeai graphs for representing higher order Go board phases and for analyzing complex capturing races and seki situations on Go boards.


Igor Kononenko (FRI), Tatjana Zrimec (FRI) in sodelavci:
Dosedanji rezultati raziskav s Kirlianovo kamero

Na seminarju bomo predstavili dosedanje rezultate raziskav bioelekričnega polja rastlin in ljudi s pomočjo Kirlianove kamere. Rusko kamero Crown-TV s programskim paketom GDV (Gas Discharge Visualization), ki omogoča učinkovito procesiranje posnetih slik z računalnikom, smo uporabili v šestih študijah: tri so relativno zaključene s pozitivnimi rezultati, v treh pa so rezultati preliminarni, vendar obetavni. Na seminarju bo opisana Kirlianova kamera, na kratko bo opisana motivacija, ki izvira iz neznanstvenih krogov: videnje aure s prostim očesom, in opisani bodo poteki in rezultati šestih študij, ki so naštete spodaj skupaj s seznamom sodelavcev:

  1. Korone lupin jabolk: študenta FRI: Aleš Doganoc in Mitja Krebelj ter dr. Marjan Simčič z Biotehnične fakultete
  2. Korone jagod grozdja: Daniel Skočaj iz Lab. za računalniški vid na FRI
  3. Vpliv različnih majic na korone prstov ljudi: Aleksander Sadikov, diplomant FRI
  4. Vpliv tehnik (dihanja itd) "Umetnost življenja" na počutje ljudi in njihove korone: Andrej Trampuž, absolvent psihologije
  5. Vpliv ovulacijskega ciklusa na korone prstov žensk: študenta FRI: Katarina Mele in Tadej Milharčič ter dr. Barbara Rojnik, dr. med - specialistka ginekologinja
  6. Relacija med diagnostiko bioterapevta in koronami prstov ljudi: mag. Bor Prihavec iz Lab. za računalniški vid na FRI, Matjaž Bevk, absolvent FRI in bioterapevt Slobodan Stanojevi} ter študenti FRI Mitja Golouh, Damjan Vavpotič, Damjan Kovač in Adamlje Peter.


Sven Erik Jorgensen (Royal Danish School of Pharmacy, Copenhagen, Denmark):
Application of Thermodynamics to Describe the Development of Ecosystems

The lecture will present different thermodynamic approaches that have been used during the last 1-2 decades to describe the development of ecosystems. It will be demonstrated that they are all consistent at least if we distinguish between ecosystems under development and mature ecosystems. It will be shown that we today can set up a very consistent description of ecosystem development by use of the concept of exergy (free energy of the system relatively to a reference system). It will be discussed how these thermodynamic concepts can be used as ecological indicators to assess the health of an ecosystem.


Aleks Jakulin (FRI):
Vizualizacija

Smisel vizualizacije je omogočiti človeku razumevanje zapletenih konceptov, na primer povezanosti, količin ter prostorov v višjih dimenzijah, tako da se ti projecirajo v barvne slike in animacije v nižjih dimenzijah. Ljudje so pri vizualnem razmišljanju zelo učinkoviti. Večina tehnik za izboljševanje pomnenja temelji ravno na vizualizaciji, na primer mnemonika. Primerjave možganskih posnetkov (PET) ljudi pri opravljanju nekaterih nalog kazejo, da bolj uspešni več uporabljajo vizualne opne (cortex) ter manj druge dele možganov.

Cilj tega seminarja je prikazati pomen vizualizacije pri analizi podatkov. Predstavitev bo temeljila na več ilustrativnih primerih in bo predvsem praktična in ne tehnična ali teoretična. Vizualizacija je komplementarna strojnemu učenju, zato se bomo tudi posebej posvetili primerjavi teh dveh področij: predvsem bomo izpostavili prednosti vizualnega opisa v primerjavi z uporabo logičnih ali lingvističnih opisov hipotez.

Naslovi programov, predstavljenih na seminarju
ViSta: http://forrest.psych.unc.edu/research/index.html
Data Explorer: http://www.research.ibm.com/dx/
Vis5D: http://www.ssec.wisc.edu/~billh/vis5d.html


Vladislav Rajkovič (FOV in IJS):
Nekateri izzivi inteligentnih sistemov za pomoč pri odločanju

Naraščajoča kompleksnost sodobnega sveta in s tem odločitev v njem spravlja snovalce sistemov za pomoč pri odločanju v nekakšen obup. Nemoč človeka in težave s sodobno informacijsko in komunikacijsko tehnologijo gre iskati tudi v naši miselnosti, saj se zdi, de ne moremo razumeti, da je bolje imeti približno prav kot pa popolnoma narobe. V podporo taki spremenjeni miselnosti lahko v managementu odločitvenega znanja s pridom uporabimo nekatere principe inteligentnih sistemov. Ogledali si bomo nekatera tuja in naša domača prizadevanja na tem področju.


Ljupčo Todorovski, Sašo Džeroski (IJS):
Meta Level Classification Trees

Different methods have been developed and used recently for improving predictive performance by combining multiple learned models. These models can be generated using a single learning algorithm on different versions of the training data (as in bagging and boosting) or using different learning algorithms on the same training data. The predictions of the models can be combined using several methods, which can be clustered in three general combining paradigms: voting, stacking and cascading.

In the presented work, we use stacking to combine the classifications of two well-known classification algorithms: the tree-learning algorithm C4.5 and the rule-learning algorithm CN2. The classifications are combined using meta-level classification trees (MLCT). A MLCT predicts which classifier should be used for the classification of an example, instead of predicting the class value itself. For the purpose of this study, the C4.5 algorithm was adapted for learning MLCT. The presented method has been empirically evaluated with experiments in 27 domains: 24 from the UCI ML repository and 3 water quality domains. Several hypothesis concerning the improvement of classification performance have been tested and approved.