Aleš Tavčar (2009) Pathology in minimin search. EngD thesis.
Game playing is one of the first areas of artificial intelligence studied by AI researchers. The evolution of adequate algorithms and heuristics have brought us to a point where computer players are able to compete with the best human players. This is achieved above all by algorithms that are able to efficiently search large game trees when choosing proper moves. In game playing it is common to examine the game tree from the current position to some depth. The states at that depth are heuristically evaluated and backed-up back to the root node, where they are used to choose a move. From this description we may surmise that deeper search gives better results than shallow, which can also be percieved in practice. However, mathematical analyses have shown that under certain conditions the opposite happens: deeper search gives worse decisions. This phenomenon was termed search pathology. In this thesis we deal with single-agent pathological models and we examine factors influencing the pathology. We have examined the properties that affect the occurrence of the pathology in synthetic search trees. The most important factors are the branching of the game tree, the dependence between nearby nodes and the number of different node values. Our main goal was to explicate their influence on the occurrence of the pathology. Next, we combined independent and dependent trees. That way we can construct partially dependent trees and analyze the pathology in such trees. Finally, we concluded that all constructed models behave similarly: the pathology is strengthened by increasing tree branching and is weakened by increasing the dependence between nearby no
Actions (login required)