Domain-Dependent Single-Agent Search Enhancements

Abstract

AI res ea rch has developed an extensive collection of methods to so lv e state-space probl ems. Using the challenging domain of Sokoban, this p ape r studies the e ffec t of search enh a ncem ents on program performance. We show that th e c urrent sta te of the art in AI genera ll y requir es a large programming and research effort int o dom ain-depe nd e nt m et hods to solve even moderately comp lex problems in such difficult domain s. The application of dom ain-sp ec ifi c kn ow ledge to exploit pr ope rti es of the search space can r es ult in large re du ct ions in th e size of the sea rch tree, often several orders of magnitude per search enh anceme nt . Understanding the e ffect of these e nhan cements on the search leads to a n ew taxonomy of sea rch e nhan cements , and a new fram ework for developing single-agent search applications. This is used to illustrate the large gap betw ee n what is portrayed in the liter at ure versus wha t is n ee ded in practice.