Generalized model for rational game tree search

Decision-theoretic meta-reasoning is a well known scheme for controlling search that has been shown to be advantageous in numerous domains, including real-time planning and acting, and game-tree search. Although in numerous adversarial games, such as chess, brute-force search currently emerges as the best contender, there is still scope for planning in some situations. In order to take advantage of both schemes, we merge the planning and exhaustive search schemes through meta-reasoning. Approximate value of information is used to decide which of the types of computation operator to apply, and where. This is done by generalizing the best play for imperfect player (BPIP) search control model of E. Baum and W. Smith, (1995) to allow for planning steps, as well as game-tree search steps. A rudimentary system employing these ideas for chess was implemented, and preliminary empirical results are promising.