Next: Evaluation using a logical
Up: Evaluation of the game
In constructing the best play for an agent, the game tree from which the play is derived must be constructed using the plan rules drawn from its belief set. Therefore, at each call to minimax, a new game tree should be constructed that is based at the level of the planning agent. Unfortunately, this is computationally expensive since there are many calls to minimax in a typical game tree. In fact game tree construction is much more expensive than evaluation, and it was found in implementation that recomputation of the game tree is not feasible on an ordinary computer. For the current design, the system only constructs one game tree. This is acceptable since in all of the examples given in this thesis, beliefs about plan rules are mutual - it is only beliefs about the domain state that differ.