Seminarji Slovenskega drustva za umetno inteligenco 1996

9.1.1996
Claude Sammut (School of Computer Science and Engineering, University of New South Wales, Sydney, Australia): An ILP Workbench
30.1.1996
Igor Kononenko, Marko Robnik-Sikonja (FRI): Nekratkovidno ocenjevanje atributov v regresiji
6.2.1996
Branko Ster: Prognostika koronarne bolezni z avtomatskim ucenjem
19.3.1996
Jus Kocijan (FE): Primer identifikacije modela sistema v delovni tocki
25.3.1996
Ruzena Bajcsy (University of Pennsylvania): Descriptive and Prescriptive Language for Mobility Task: Are they different?
7.5.1996
Bogdan Filipic (Institut Jozef Stefan in Fakulteta za strojnistvo, Ljubljana), Mihael Junkar (Fakulteta za strojnistvo, Ljubljana): Od predprogramiranih strojev k inteligentnim obdelovalnim sistemom
14.5.1996
Bern Martens (Department of Computer Science, Katholieke Universiteit Leuven, Belgium): Tutorial on Logic Program Specialisation
21.5.1996
Marko Bohanec (IJS): Ptah: Sistem za podporo terapevtskih aktivnosti pri hospitalnih infekcijah
1.10.1996
Darko Zupanic (IJS): Uporaba sistemov za avtomatsko ucenje pri izboru materialnega para in dolocanju rezalnih parametrov struzenja
3.10.1996
Rainer Decker (GenSym GmbH, Muenchen, Germany): G2: A Comprehensive Environment for Creating Intelligent Production Management Systems
8.10.1996
Igor Zelic (Infonet, d.o.o., Kranj): Avtomatsko ucenje pri diagnosticiranju sportnih poskodb
15.10.1996
Matjaz Kukar, Ciril Groselj: Aplikacija strojnega ucenja pri diagnosticiranju ishemicne bolezni srca
29.10.1996
Matjaz Gams, Aram Karalic (IJS): EMA: Inteligentni agent za zaposlovanje
5.11.1996
Marek Perkowski (Portland State University, Department of Electrical Engineering, USA): What can Logic Synthesis Offer to Machine Learning?
12.11.1996
Zdravko Pecar (Univerza v Ljubljani, Visoka upravna sola): Aplikacija metod in orodij umetne inteligence na podrocju managerskega odlocanja
26.11.1996
France Mihelic, Simon Dobrisek, Ivo Ipsic, Karmen Pepelnjak, Jerneja Gros (Univerza v Ljubljani, Fakulteta za elektrotehniko): Predstavitev del v Laboratoriju za umetno zaznavanje
3.12.1996
Rajko Mahkovic (FRI): Nacrtovanje gladkih poti za mobilne robote
10.12.1996
Damjan Bojadziev (IJS): Formalna refleksija v umetni inteligenci

Claude Sammut (School of Computer Science and Engineering, University of New South Wales, Sydney, Australia): An ILP Workbench
Data mining has become a popular research topic in recent years. This talk describes a project whose aim is to investigate the usefulness of Inductive Logic Programming for data mining. We focus on software that is being developed as a test bed for the investigation.

The workbench is centred around a conventional Prolog interpreter. However, the interpreter as an attached database indexing scheme to allow rapid access to large relations. In addition, a library built-in predicates provides the user with a number of machine learning tools for exploring the data contained in the data base. The ML tools include attribute/value learning algorithms as well as ILP algorithms.


Igor Kononenko, Marko Robnik-Sikonja (FRI): Nekratkovidno ocenjevanje atributov v regresiji
Ocenjevanje atributov je ena kljucnih tock v strojnem ucenju. Vecina algoritmov pri ocenjevanju kvalitete atributov predpostavlja njihovo neodvisnost, zato so neprimerni v problemih, kjer nastopajo med atributi mocne soodvisnosti. Na podlagi analize algoritma ReliefF, ki je uspesen pri klasifikacijskih problemih iste vrste, sva razvila RReliefF (regresijski ReliefF), ki uspesno odkriva odvisnosti med atributi v regresijskih domenah. Predstavila bova algoritem RReliefF in poiskuse, s katerimi sva preverjala njegovo ucinkovitost pri ocenjevanju kvalitete atributov in ucenju regresijskih dreves.
Branko Ster : Prognostika koronarne bolezni z avtomatskim ucenjem
Avtor bo prikazal rezultate uporabe sistemov: LFC, Assistant-I, Assistant-R, naive Bayes, semi-naive Bayes, backpopagation, backpropagation with weight elimination, CART, Fisherjeva diskriminantna, linearna diskriminanta, kvadratna diskriminanta, K-NN, LVQ in RBFL (radialne funkcije) na problemu prognostike koronarne bolezni. Rezultati kazejo na to, da sedanji atributi ne nudijo nobene koristne informacije za dani prognosticni problem. Poleg tega bo avtor primerjal rezultate zgoraj nastetih sistemov na nekaterih "benchmark" podatkovnih bazah.

Jus Kocijan (FE): Primer identifikacije modela sistema v delovni tocki
Prikazal bom primer prakticne identifikacije dinamicnega sistema s posebnim poudarkom na izbiri vhodnega signala, casa vzorcenja, predhodni obdelavi signalov, izbiri modela in preskusu njegove veljavnosti. Namen prispevka je osvetliti nekatera vprasanja, ki so pri opisu identifikacijskih metod velikokrat postavljena nekoliko v ozadje, so pa nujna v praksi identifikacije dinamicnih sistemov. Posebej je predstavljena nujnost po iterativnosti uveljavljenih postopkov, ki lahko privedejo do najboljsega moznega modela po zadanih kriterijih.
Ruzena Bajcsy (University of Pennsylvania): Descriptive and Prescriptive Language for Mobility Task: Are they different?
In this presentation we wish to address the following problem: What is the necessary (and possibily sufficient) VOCABULARY that will DESCRIBE scenes of moving vehicles (traffic scenes) as observed via a TV camera or cameras and at the same time will serve as a TASK DESCRIPTION LANGUAGE to command and Autonomous Land vehicle in their driving task. In other words, will the same vocabulary be sufficient to PRESCRIBE the Task and the necessary information about the vehicle and its environment interaction in order to accomplish the driving task. We view this problem as fundamental to the REPRESENTATION of the necessary dynamical knowledge/information about the agent (i.e., the car and its driver), the environment (the road and all the objects on the road, including other cars), the task and their interactions. The issue of "how to represent" in our interpretation is how to partition the continuous dynamical process of the agent/environment/task interaction into discrete (symbolic) states, events, behaviours and places. While the Descriptive part is the classical SIGNAL to SYMBOL transformation, in our case going from visual observations into symbolic decsriptions, the Prescritpive part is the reverse, that is going form SYMBOLS to SIGNAL, that is translating the Symbolic expressions into continuous behaviours.
Bogdan Filipic (Institut Jozef Stefan in Fakulteta za strojnistvo, Ljubljana), Mihael Junkar (Fakulteta za strojnistvo, Ljubljana): Od predprogramiranih strojev k inteligentnim obdelovalnim sistemom
Za danasnjo tehnologijo obdelave materialov sta znacilni visoka stopnja avtomatizacije in ozka specializiranost. Predprogramirani, numericno krmiljeni obdelovalni stroji so izredno zmogljivi, a hkrati zaprti in tezko prilagodljivi na spremembe v delovnem okolju. Ob njih ostaja neizkorisceno bogastvo operaterjevega znanja in izkusenj. Vecja prilagodljivost, zmoznost ucenja in izpopolnjevanje v casu obratovanja so zato izzivi za nacrtovalce novih generacij obdelovalnih sistemov. Uporaba metod strojnega ucenja nakazuje pot k tovrstnim izboljsavam. V predavanju bo predstavljeno ucenje nadzora procesa brusenja, izbire brusilnega orodja in vodenja elektroerozijske obdelave kovin. Delo bo osvetljeno tako z metodoloskega kot z uporabniskega vidika.
Bern Martens (Department of Computer Science, Katholieke Universiteit Leuven, Belgium): Tutorial on Logic Program Specialisation
We give a broad and general introduction to program specialisation, including motivating examples for partial deduction, and related forms of specialisation, such as functor elimination and generation of most specific programs. We then provide some insight into the foundations of partial deduction, including correctness and completeness conditions from the well-known paper by lloyd and Shepherdson. We illustrate ways in which they can be ensured.

Next, we elaborate issues related to the on-line control of partial deduction. We motivate and illustrate the distinction between local and global control in partial deduction algorithms, point out some trade-offs, and briefly sketch approaches based on termination analyses, on characteristic trees, and on a combination of both.

We point out relations between partial deduction and abstract interpretation as well as unfold/fold transformations. Finally, we discuss some specific problems related to specialisation of meta-programs and indicate some solutions. We end with the presentation of some open problems.


Marko Bohanec (IJS): Ptah: Sistem za podporo terapevtskih aktivnosti pri hospitalnih infekcijah
Na seminarju bo predstavljen racunalniski program Ptah za podporo terapevtskih aktivnosti pri hospitalnih infekcijah. Program smo razvili v okviru razvojnega projekta "Racunalniski informacijski sistem za podporo terapevtskih aktivnosti v zdravstvu", ki poteka v sodelovanju s podjetjem Infonet, d.o.o., Kranj in Splosno bolnisnico Jesenice.

Osnovo za delovanje programa predstavlja zbirka podatkov o hospitalnih infekcijah, iz katerih gradi casovne vrste vzorcev rezistentnosti bakterij. Uporabniku (zdravniku) ponuja pet metod za analizo ucinkovitosti antibiotikov, odpornosti bakterij na antibiotike in identifikacijo hisnih bakterij. Rezultati analiz, ki so vecinoma predstavljeni graficno, pomagajo zdravniku pri odlocanju in izbiri najprimernejse terapije.

Ptah je realiziran v okolju MS Windows in je v fazi prakticnega uvajanja.


Darko Zupanic (IJS): Uporaba sistemov za avtomatsko ucenje pri izboru materialnega para in dolocanju rezalnih parametrov struzenja
Nacrtovanje rezalnih parametrov je zaradi empiricno dolocenih modelov algoritmsko dokaj neobvladljivo. Znanje o domeni je vsebovano v mnozici razlicnih katalogov in smernicah posameznih proizvajalcev. Za vsak materialni par je potrebno izvajati drage in dolgotrajne meritve, ki pa zaradi specificnih pogojev ne dajo splosno uporabnih pravil. Raziskovalci so pri resevanju teh problemov uporabljali statisticna orodja za analizo podatkov, ki so dala le delne rezultate. Metode avtomatskega ucenja so odprle dodatne smeri raziskav, saj dopuscajo moznost odkrivanja kompleksnejsih relacij. Te relacije lahko predstavimo kot pravila, ki jih nato uporabljamo v procesu avtomatskega nacrtovanja tehnologije obdelave.
Rainer Decker (GenSym GmbH, Muenchen, Germany): G2: A Comprehensive Environment for Creating Intelligent Production Management Systems
The presentation covers two sections. The first section overviews G2 technologies and how these technologies work together. The second section covers the overall approach to delivering intelligent applications with G2. G2 integrates a complete set of technologies for rapidly building and delivering intelligent applications. The core of G2 technologies includes its concurrent real-time execution together with object-oriented design.

Integrated into G2 core are its interactive graphics, rules, procedures and methods, structured natural language and dynamic modeling and simulation capabilities. These capabilities directly support development. A next layer for technologies are those needed for deployment such as client/server networking and connectivity with on-line data. G2 is designed especially for real-time operation. G2 powerful real-time capabilities have evolved over many years of on-line application development.

The heart of G2 real-time operation is a scheduling engine for concurrently executing rules, procedures, models, and other tasks. Typically many lines of reasoning, models, and other activities are occurring concurrently. The engine cycles according to a clock "tick" that can be specified at a millisecond level. Everything has a priority so that critical items can take action first.

Object-oriented design is the basis of development in G2. Built around the concept of objects, object-oriented design offers highly reusable code and an application structure that is much more intuitive to understand than those built with conventional programming tools. G2 rules, procedures, and methods typically represent a large percentage of the knowledge contained in an application. The ability to create generic rules, procedures and methods that apply across entire classes of objects shortens the development time and makes it easier to read and maintain the knowledge in the applications.

G2 is designed for connecting to all kind of on-line real-time data. Gensym offers off-the-shelf bridges for linking to common commercial systems as well as data interface toolkits for building connections to data sources.

Gensym products apply intelligence to production management systems to extend its ability of solving complex problems. G2 performs functions that would be prohibitively expensive to implement in conventional process control systems or to program conventionally in supervisory control systems.

In process control and optimization, intelligent systems apply advanced technologies to improve operational efficiencies, even across an entire enterprise. These technologies are complimentary to the proportional control methods emphasized by DCS and the logical methods that dominate in PLC. They include algorhitmic techniques using standard mathematical functions such as linear programming, heuristic techniques that emphasized experience-based rules, artificial neural networks which use process data histories to form current estimates and conclusions, and many other techniques like constraint-directed reasoning, fuzzy logic, and genetic algorithms.

In monitoring and diagnosis, intelligent systems improve reliability of data, apply predictive maintenance (meaning that they detect problems earlier than is possible with conventional systems), and help pinpoint problem root causes so that they can be more readily corrected. In this way, they compliment conventional control systems. For the most part, conventional systems are limited to simple comparisons of signal values against alarm limits, and they offer little assistance to the operator in determining why a particular alarm condition exists. Intelligent systems are also finding increased use in off-line applications, particularly in conjunction with dynamic simulators used for operator training and in plant design.

The attractiveness of G2 comes down to its completeness in combining several powerful technologies for developing an deploying intelligent systems. Customers report significant development time reductions and corresponding time to installation or to market. In other words, G2 provides a comprehensive infrastructure for application development.


Igor Zelic (Infonet, d.o.o., Kranj): Avtomatsko ucenje pri diagnosticiranju sportnih poskodb
Na seminarju bo predstavljena aplikacija avtomatskega ucenja diagnosticiranja sportnih poskodb. Sistem je bil razvit v sodelovanju z dr. Vugo iz Centra za medicino sporta pri UKC. Izvedene so bile primerjave naslednjih algoritmov: Bayesov klasifikator, Delno naivni Bayes, Asistent-I, Asistent-R in Asistent-R2. Primerjali smo tako klasifikacijsko tocnost in informacijsko vsebino odgovorov kot zmoznost in razumljivost razlage posameznih algoritmov. Zaradi majhnega stevila ucnih primerov smo algoritmom za avtomatsko ucenje dodali tudi diagnosticna pravila. Razvita lupina ekspertnega sistema je razvita v okolju MS-Windows in omogoca manipulacijo in statisticno obdelavo domene, kontroliran in uporabniku prijazen vnos ucnih in testnih primerov, testiranje algoritmov ter klasifikacijo novih primerov z ustrezno graficno razlago.
Matjaz Kukar, Ciril Groselj : Aplikacija strojnega ucenja pri diagnosticiranju ishemicne bolezni srca
Ishemicna bolezen srca je eden od najpogostejsih vzrokov smrtnosti v razvitem svetu, tako da bi bili izboljsave in racionalizacija diagnosticnih postopkov zelo uporabni. Postavljanje diagnoz poteka v stirih fazah (glede na opravljene preiskave): (1) status in anamneza, (2) EKG ob mirovanju in kontrolirani obremenitvi (cikloergometrija), (3) miokardna scintigrafija in (4) koronarna angiografija. Diagnosticni proces poteka hierarhicno - naslednja faza je potrebna samo, ce rezultati prejsnje niso dovolj zanesljivi.

Zdravnikom pa se pri postavljanju diagnoze pojavlja problem - apriorno ustvarjanje mnenja o pacientu (sugestibilnost). Zato se rezultati vsake faze interpretirajo individualno, veljavni pa so samo rezultati zadnje faze.

Algoritmi za strojno ucenje niso podvrzeni subjektivnim vplivom (sugestibilnost), zato bi lahko upostevali rezultate vseh ze opravljenih preiskav naenkrat in s tem potencialno izboljsali postopek diagnosticiranja. Prvi poskusi so ze dali obetavne rezultate, ki jih bomo predstavili na seminarju.


Matjaz Gams, Aram Karalic (IJS): EMA: Inteligentni agent za zaposlovanje
Septembra 1996 se je koncal inovacijski projekt Integrirani informacijski sistem za izboljsanje zaposlovanja v Sloveniji, najvecji inovacijski projekt umetne inteligence doslej, polovico financiran s strani narocnikov. Med stirimi deli sta dva direktno navezana na Odsek za inteligentne sisteme: Katalog programske opreme in Inteligentni agent za zaposlovanje. Posebej zadnji je vec kot leto dni v redni uporabi na Internetu in je bil klican okoli 20.000-krat. Sistem ima okoli 10.000 vrstic, skupaj s podatki in besedili zasega 30M.

Ema - inteligentni(a) zaposlovalni(a) agent(ka) - je zadnji, vrhnji nivo celotnega sistema, ki poslje elektronsko posto, ko se na primer pojavi zanimivo delovno mesto pomocnika direktorja. Vsak teden sistem predstavi okoli 3.000 razpisov za delovna mesta, mozno se je vpisati kot iskalec dela, tu so stipendije itd.


Marek Perkowski (Portland State University, Department of Electrical Engineering, USA): What can Logic Synthesis Offer to Machine Learning?
The lecture will introduce a new approach to Machine Learning that uses logic synthesis methods close to those used in circuit design, and especially for Xilink FPGAs. The data to be learned are represented as Boolean or multiple-valued functions. The input variables correspond to the features of the image, and the minterms of the function to the positive or negative samples, or images classified as belonging or not to the specified category. Multi-category classification is performed using multi-output, multiple-valued functions.

An interesting property of applying logic synthesis to Machine Learning is that the functions can be very strongly unspecified, having more than 99% of don't care conditions. This calls for the entirely new approaches to the algorithm development. Logic-based approaches based on Sum-of-Products, Exor-Sum-of-Products, and Ashenhurst-Curtis Functional Decompositions will be compared with the classical approaches of Neural Nets and Decision Trees.

The method has been successfully used for Pattern Recognition, noise removal from images, Knowledge Discovery in Data Bases, and automatic creation of algorithms from data.


Zdravko Pecar (Univerza v Ljubljani, Visoka upravna sola): Aplikacija metod in orodij umetne inteligence na podrocju managerskega odlocanja
Sprejemanje odlocitev predstavlja najzahtevnejso fazo procesa managementa. Poslovne odlocitve morajo biti vse bolj hitre in natancne, vsaka odlocitev pa zahteva predhodno tudi vedno vec razlicnih proucevanj podatkov in informacij. Tukaj nastane velika nevarnost, da manager ne zeli pridobiti vseh informacij in se raje odloca po svojem instinktu, ali pa da razpolaga s preveliko kolicino informacij, ki jih zaradi pomanjkanja casa in premalo razvite tehnologije ne more obdelati tako, da bi bile maksimalno izkoriscene. Nastalo zagato bi najlazje resevali s pomocjo primerno prilagojenih informacijskih tehnologij, ki bi morale biti bolj naravnane na procese odlocanja in resevanja problemov.

Podrocje proucevanj procesov odlocanja in se posebej v kontekstu sirsih procesov resevanja problemov je relativno mlada znanstvena disciplina, ki predstavlja v kombinaciji z orodji umetne inteligence velik potencial za lazje delo managerjev. Na seminarju bodo predstavljene nekatere ideje in preliminarni rezultati proucevanja teh potencialov, ki je potekalo v okviru priprave doktorske disertacije.


France Mihelic, Simon Dobrisek, Ivo Ipsic, Karmen Pepelnjak, Jerneja Gros (Univerza v Ljubljani, Fakulteta za elektrotehniko): Predstavitev del v Laboratoriju za umetno zaznavanje

1. razpoznavanje slovenskega govora:

aplikaciji:

2. sinteza slovenskega govora:


Rajko Mahkovic (FRI): Nacrtovanje gladkih poti za mobilne robote
Malo je raziskovalnih podrocij, ki bi v zadnjem casu pritegnila tako stevilne raziskovalce, kot so jih mobilni roboti. Problemov lokalizacije, navigacije, nacrtovanja poti, preiskovanja neznanega prostora, pridobivanja avtonomnosti so se lotili tako robotiki kot racunalnikarji (prevsem AI). Na seminarju bomo predstavili del problematike in resitev na podrocju nacrtovanja poti, posebno se gladkih poti. Prikazali bomo voznjo v laboratoriju LMSA razvitega mobilnega robota.
Damjan Bojadziev (IJS): Formalna refleksija v umetni inteligenci
Formalizacija inteligence in zavesti zahteva uporabo refleksivnega sistema, ki zmore vpogled v lastno delovanje. Tak sistem je nepopoln in njegov "Goedlov operator" proizvaja neskoncno zaporedje razsiritev. Operator razsiritve pa bi lahko bil tudi sam refleksiven: z upostevanjem ucinka lastnega delovanja bi lahko preskocil dele zaporedja razsiritev. Domneva o obstoju refleksivnega operatorja razsiritve izhaja iz refleksivnosti postopka, ki proizvede Goedlov samo-referencni stavek, iz Kleenejevega izreka o rekurziji in delno iz moznosti refleksivne uporabe meta-cirkularnega interpreterja. Toda iz Turingovih rezultatov o obnasanju univerzalnega stroja sledi, da "turbo-refleksija", ki bi lahko preskakovala stopnje obicajne refleksije, dejansko ni mogoca.