Bogdan Filipič
(IJS in Univerza v Ljubljani, Fakulteta za strojnitvo):
Optimiranje proizvodnih procesov z metodami evolucijskega računanja
Metode evolucijskega računanja proučujemo v Odseku za inteligentne sisteme od leta 1991. Delo trenutno poteka v okviru raziskovalnega projekta Evolucijsko računanje v optimizaciji in identifikaciji sistemov, sodelujemo pa tudi v mednarodnem projektu EvoNet (European Network of Excellence in Evolutionary Computing). Med drugim se posvečamo razvoju evolucijskih algoritmov za reevanje optimizacijskih problemov v proizvodnih procesih. Na seminarju bosta predstavljena dva noveja primera prenosa te metodologije v prakso: časovno razporejanje procesov v obratu stiskalnic za avtomobilsko karoserijo in optimiranje procesnih parametrov pri ulivanju jekla. Prikazali bomo optimizacijska problema, način reevanja in dosedanje rezultate ter povzeli izkunje, pridobljene na obeh nalogah.
Nada Lavrač
(IJS):
Obravnava problema relevantnosti v strojnem učenju in
uporaba algoritma REDUCE
Na seminarju predstavimo problem relevantnosti v strojnem učenju. Algoritem REDUCE omogoča eliminacijo irelevantnih literalov/selektorjev v predprocesiranju učne množice. Študija, v kateri smo uporabili REDUCE kot predprocesor genetskega algoritma, pokaže uporabnost algoritma za zmanjševanje časovne kompleksnosti učenja z genetskim algoritmom ter izboljšanje rezultata učenja, merjenega s kompleksnostjo naučene hipoteze. Delo je nastalo v sodelovanju z Draganom Gambergerjem (Institut Rudjer Boškovič, Zagreb) in Petrom Turneyem (National Research Council Canada, Ottawa).
Nada Lavrač
(IJS):
Eliminacija šuma v predprocesiranju podatkov
Praktični algoritmi za strojno učenje vsebujejo mehanizme za obravnavo šumnih podatkov, npr. mehanizme za rezanje pravil in odločitvenih dreves. Na seminarju predstavimo algoritem za odkrivanje in eliminacijo šumnih primerov v predprocesiranju podatkov. Osnovni algoritem, ki je bil razvit v okviru sistema ILLM, smo razširili na obravnavo večrazrednih problemov ter ga evaluirali v več domenah: dveh šahovskih končnicah z umetno dodanim šumom in desetih realnih medicinskih domenah. Delo je nastalo v sodelovanju z D. Gambergerjem (Institut Rudjer Boškovič, Zagreb) in S. Džeroskim (IJS).
Marko Robnik-Šikonja
(FRI):
Ocenjevanje atributov, konstruktivna indukcija, linearni modeli in rezanje
pri učenju regresijskih dreves
Pogledali si bomo analizo nekaterih ključnih elementov v znanih sistemih za učenje regresijskih dreves ter poskuse njihovih dopolnitev in izboljšav.
Pri strojnem učenju se je za enega ključnih elementov izkazala hevristična ocena kvalitete atributov. Prikazanih bo nekaj novejših rezultatov z regresijsko varianto algoritma Relief, ki je sposoben zaznati odvisnosti med atributi.
Pri klasifikaciji in v induktivnem logičnem programiranju se je na nekaterih vrstah problemov pokazalo, da obstoječi atributi in relacije ne zadoščajo za razumljiv opis danega koncepta. Tipično se problemi takšne vrste rešujejo z avtomatskim ali ročnim dodajanjem novih, vmesnih konceptov. Ta pristop, imenovan konstruktivna indukcija, smo poskusili tudi pri učenju regresijskih dreves. Uporabili smo operatorje konjunkcije, seštevanja in množenja. V duhu principa najkrajše dolžine opisa (MDL) smo izpeljali kodiranje ocene kvalitete konstruktov za RReliefF in MSE.
Princip MDL smo uporabili pri gradnji linearnih modelov v listih, za kar smo razvili kodiranje modelov, ter ga testirali z nekaj optimizacijskimi metodami.
Kodiranje konstruktov in modelov smo uporabili pri rezanju dreves po principu MDL. Novo rezanje smo primerjali z uveljavljeno metodo rezanja z m-oceno verjetnosti.
Novosti smo testirali v za ta namen razvitem učnem sistemu na več množicah umetnih in realnih podatkov.
Luc De Raedt and Hendrik Blockeel
(Katholieke Universiteit Leuven, Belgium):
Top-down induction of first order logical decision trees:
classification, regression and clustering
A first order approach to learning decision trees will be presented. The approach employs the learning from interpretations setting of Inductive Logic Programming. Two systems will be presented that learn in this setting:
Matjaž Gams
(IJS):
Inteligentni sistemi in inteligentni agenti
Paule Ecimovic
(Slovenska znanstvena fundacija;
Filozofska fakulteta, Oddelek za logiko):
Grenander's geometries of knowledge as an approach to KDD
The focus of this presentation will be the application of Ulf Grenander's system of pattern-theoretic knowledge representation, geometries of knowledge, in the process of knowledge discovery in databases (KDD). The KDD process will be presented, briefly, in the Fayyad-Piatetsky-Shapiro-Smyth scheme (Fayyad et al, KDD-96), integrating data mining as the key step in the KDD process. The concept of meta-level iteration (feeding back the knowledge produced into the next run of the KDD process) will be defined for this scheme, and its use for knowledge justification and clensing will be considered in passing. Finally, the geometries of knowledge approach will be formulated for the task of "detecting" volcanoes on Venus (Smyth et al, 1996) from 100 Mb worth of optical data gathered by the Magellan probe in 1992-94. The geometries of knowledge approach to volcano "detection" will be compared to the volcano detection algorithm employed by Smyth et al.
Philip E. Agre
(Department of Communication, University of California, San Diego):
Embodiment and cognitive architecture
I will review the traditional view of cognitive architecture, and then I will describe some changes that seem necessary as we investigate an embodied agent interacting with the physical world. Most of my discussion will be drawn from my book "Computational Theories of Interaction and Agency".
An emerging movement in artificial intelligence research has explored computational theories of agents' interactions with their environments. This research has made clear that many historically important ideas about computation are not well-suited to the design of agents with bodies, or to the analysis of these agents' embodied activities. This presentation will review some of the difficulties and describe some of the concepts that are guiding the new research, as well as the increasing dialog between AI research and research in fields as disparate as phenomenology and physics.
Beth Meyer
(Human Attention and Performance Laboratory, Georgia Institute of
Technology, Atlanta):
Adaptive user interfaces:
Principles of usability for younger and older users
Some adaptive or intelligent user interfaces have made software easier to use, while others have actually been harder to use than a normal system. This talk will discuss the differences between adaptive interfaces that work and those that don't. It will present design guidelines for adaptive interfaces based on basic psychology research. The talk will also include a brief summary of our laboratory's research on aging and the use of technology. Finally, I will present new research that shows that people deal with changes in their environment differently as they get older.
Donald Michie
(Professor Emeritus of Machine Intelligence, University of Edinburgh,
UK
Associate Member of the Jožef Stefan Institute, Ljubljana,
Slovenia):
Alan Turing and the 'Child-Machine' Approach
In the 7th and last section of his 1950 paper in Mind, Turing made the following proposal:
In the process of trying to imitate an adult human mind we are bound to
think a good deal about the process which has brought it to the state that
it is in. We may notice three components.
Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain.
Turing's career will be traced from its inception in 1936, leading to its logical culmination in the above conclusion. Options for today's continuing quest for machine intelligence will be analysed.
Claude Sammut
(Department of Artificial Intelligence,
School of Computer Science and Engineering,
University of NSW, Sydney, Australia):
Extracting hidden context
Concept drift due to hidden changes in context complicates learning in many domains including financial prediction, medical diagnosis, and network performance. Existing machine learning approaches to this problem use an incremental learning, on-line paradigm. Batch, off-line learners tend to be ineffective in domains with hidden changes in context as they assume that the training set is homogeneous.
We present an off-line meta-learning approach for the identification of hidden context. The new approach uses an existing batch learner and attempts both to identify likely context boundaries and to combine similar hidden contexts in a process called contextual clustering. The resulting data sets are then be used to produce context specific, locally stable concepts. Several algorithms for the approach are presented and evaluated. The approach is also evaluated on a complex learning task.
Richard L. Amoroso
(The International Noetic University, Physics Lab, Berkeley, USA):
Engineering a conscious quantum computer
Feynman found nothing in the laws of physics to oppose realization of quantum computers (QC), Although the nonlinear trigger of synaptic vessicle release and evidence of quantum information processing at the microtubule suggest that the brain is a QC; this is not sufficient for consciousness. Mind <> Computer because agency has an acausality beyond any algorithmic or heuristic programming. If 'res cogitans' is not immaterial as promulgated in DesCartes mind - body dichotomy mind is accessible to the laws of physics. According to this model the most viable direction for AI is the extracellular containment of natural intelligence in the core of a specific class of quantum computer simulating quantum brain dynamics. The quantum computer discussed will be Molecular Electronic - nanoscale heterosoric crystals utilizing optical holographic multi-mesh hypercube interconnects.
Karl H. Pribram
(Professor Emeritus, Stanford University):
The Brain, the Me and the I
General systems theory is based on the finding that often collectives of different scales can be shown to operate according to the same - or at least very similar - principles of organization. This is an intriguing finding but does not tell us how - the process by which - such organizations come about.
In order to determine the "how" of processing, it becomes necessary to identify the transfer functions that make it possible for operations at one scale to influence those at the adjacent higher and lower order scales. A "scale" or "level" is defined as a "description" of the organization of the elements in a system that is simpler than a description of the elenents themselves. The recognition of multiple levels of organization in a system, is not, as Robert Hinde (1992: 1,019) rightly points out, an argument for reductionism because each level "must be thought of not as an entity but rather in terms of processes continually influenced by the dialectical relations between levels." Such an approach, which requires scientists to 'cross and recross' the boundaries between levels (Hinde's phrase, quoted in Bateson, 1991: 14), thus demands integration between adjoining disciplines.
It thus becomes imperative to delineate scales describing the operations of the brains of the persons composing their minds. This lecture attempts such a delineation.
Bateson, P. (1991). Levels and processes. In P. Bateson (Ed.) The Development and Integration of Behavior: Essays in Honour of Robert Zinde (pp. 3-17). Cambridge, England: Cambridge University Press.
Hinde, R.A. (1992) Developmental psychology in the context of other behavioral sciences. Developmental Psychology, 28 (6), pp. 1018-1029.
Dorian Šuc
(FRI):
Rekonstrukcija veščine vodenja v obliki LQ kontrolerjev s podcilji
Vodenje dinamičnega sistema, kot je žerjav ali letalo, zahteva izurjenega operaterja. Ta veščina vodenja je lahko osnova za avtomatski kontroler, vendar jo je navadno zelo težko rekonstruirati z introspekcijo. Ideja vedenjskega kloniranja je rekonstrukcija veščine vodenja s strojnim učenjem iz posnetkov operaterjevega vodenja. Osnovni pristop k vedenjskemu kloniranju pa skriva v sebi številne pomanjkljivosti.
Na seminarju bo predstavljen pristop k vedenjskemu kloniranju, ki za razliko od klasičnega pristopa upošteva tudi dinamiko sistema vodenja ter omogoča identifikacijo operaterjevih podciljev. Ta pristop uporablja nekatere rezultate klasične teorije vodenja (LQ kontrolerji), inducirani kontrolerji pa so robustni in omogočajo razlago operaterjeve veščine vodenja.