Chapter 12 - Adversary Search

Abstract

This chapter studies heuristic search in game graphs. For two-player games, tree search from the current node is performed and endgame databases are built. The chapter discusses different refinement strategies to improve the exploration for forward search and suggests a symbolic classification algorithm. Adversaries introduce an element of uncertainty into the search process. Optimal strategies result in perfect play. In most settings, the players can take actions alternately and independently. In contrast to the deterministic search models, the outcome of an executed action in a state is not unique. Each applicable action may spawn several successors. There are many reasons for such uncertainty: randomness that is in the real world, the lack of knowledge for modeling the real world precisely, the dynamic change in the environment that one cannot control, sensors and actuators that are imprecise, and so on. Solutions to nondeterministic and probabilistic search tasks are no longer a sequence of actions but mappings from state to actions. As opposed to linear solution sequences, adversary search requires state space traversal to return solution policies in the form of a tree or a graph. The policy is often implicitly represented in the form of a value function that assigns a value to each of the states. For the deterministic setting, the value function takes the role of the heuristic that is gradually improved to the goal distance. This links the solution process for adversary search problems to the ones presented for real-time search, where the value function for deterministic search models is improved over time.