Experience-Based Creativity

A chess program that uses its transposition table to learn from experience. Explaining temporal diierences to create useful concepts for evaluating states. 35 { What happens when Morph is trained on these games? What if the games are presented in reverse order? Can we also explain Bobby Seltzer's development in terms of learned patterns? Can his \creative" moves be explained in terms of past experiences? { We have a similar dataset of Bobby Fischer's games? Which Bobby does Morph learn best from: Fischer or Seltzer? A very long term goal is to get Morph to be strong enough to emulate a match between these two (or Fischer and the current World Champion). Of critical importance is the determination of more compelling domains beyond chess to which these methods can be applied. One such domain would seem to be organic synthesis 23,51]: through experience a system could learn which types of molecules are more easily or cheaply made and guide synthesis pathways in this direction. Similar ideas may also apply to automatic theorem proving. Finally, we believe that the \psychology of Morph" is another direction that is worth pursuing. For example, we have witnessed signs of depression at times: many patterns are evaluated negatively, or Morph is afraid to try new things or to try things that failed early in its training. Likewise, from time to time Morph plays extremely aggressively and sometimes recklessly. Where does this tendency come from? Perhaps, it is through understanding the mathematical principles behind the eeect of experience on decision-making and future experience that better therapies can be developed. Can Morph's creativity be \encouraged?". 7 Acknowledgements We are indebted to those researchers whose methods have been adapted for Morph, Morph's programmers, the encouragement of friends and family, NSF grant IRI-8921291 and to experience itself. 34 guide search in tactical situations. Paradise was able to nd combinations as deep as 19-ply. It made liberal use of planning knowledge in the form of a rich set of primitives for reasoning and thus can be characterized as a \semantic approach." This diierence plus the use of search to check plans and the restriction to tactical positions distinguish it from Morph. Also, Paradise is not a learning program: patterns and planning knowledge are supplied by the programmer. Epstein's Hoyle system 7] also applies a semantic approach but to multiple simultaneous game domains. 6 Conclusions and Directions As we nish preparing …