Pathology in heuristic search algorithms

The thesis deals with minimax and single-agent search. Practice shows that in both cases deeper search results in better decisions. Mathematical analyses by previous researchers, however, have shown that the opposite is true: that minimaxing amplifies the error of the heuristic evaluations and hence deeper search gives worse decisions. This phenomenon has been termed search pathology. In the thesis we use a minimax model with real numbers as both true and heuristic position values to show that proper modeling of the heuristic error is enough to prevent the pathology. We proceed to examine how branching factor, dependence between nearby positions and the number of possible position values affect the pathology. We also show why minimax search is actually beneficial. The pathology in single-agent search is explained on synthetic search trees and path-finding on maps from computer games. We finally explain why single-agent search is beneficial