Year: 2006
Type: dissertation
Abstract: This thesis investigates rule extraction from recurrent neuralnet"'orks, which tak('~ t he form of autOlnated construction of models of an underlying network.T~'pically the models are expressed as finite state machines and they should Inimic t he network while being more intelligible.It is argued that rule extraction allows a deeper and more general form of analysis than other, Inore or less ad hoc, Inet hods which are typically applied after the training of the recurrent networks.The first part of this thesis reviews and analyses the development of related techniques.The s{'('owl part presents a novel algorithm, the Crystallizing SlIbstochastic SCfjucntial Alachine Extractor (CrySSMEx), which efficiently generates a sequence of increasingly refilled stochastic finite state Inodels of an underlying s,•stein.:; ovel features of CrySSMEx include, for example, freedOln froill paraIneters.detenllinistic extraction, a hierarchical vector quantizer.and a stochastic finite state Illodel which can be constructed also when sonle data is lllissing.Experiments show that CrySSMEx is, cOlllpared to other Inethods.applicable to a wider range of problellls (such as high-diIllensional or chaotic clynaIllic systems).Finally, the field is discussed fronl a lnore theoretical perspective in terms of scientific Inethoclology targeted at silnulated systeIn~.It is suggested that a rule extractor (or Empirical ~Iachille) can actively select data frOln the Sy'steIll it is set to IHodel by continuously targeting the weakest point of its currently strongest Illodel.These alltOlllated experilnenters can, in turn, be nwcle part of a frarnework (or Popperian ~Iachine) ill which theories about populatiolls of systeins are generated and tested in order to est ablish falsifiable SLl telnents.These stateinents should have a high eillpirical content and thus concisely describe einergent, and previously unknown, properties of the systeIlls. Technical reportsJacobsson, H. and Ziemke, T. (2003).Reducing Complexity of Rule Extraction from Prediction RNN s trough Domain Interaction 2 , Tech.report no.HS
Author: Henrik Jacobsson
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1568782853