Safe Opponent-Exploitation Subgame Refinement
Search algorithms have been playing a vital role in the success of superhuman AI in both perfect information and imperfect information games. Specifically, search algorithms can generate a refinement of Nash equilibrium (NE) approximation in games such as Texas hold'em with theoretical guarantees. However, when confronted with opponents of limited rationality, an NE strategy tends to be overly conservative, because it prefers to achieve its low exploitability rather than actively exploiting the weakness of opponents. In this paper, we investigate the dilemma of safety and opponent exploitation. We present a new real-time search framework that smoothly interpolates between the two extremes of strategy search, hence unifying safe search and opponent exploitation. We provide our new strategy with a theoretically upper-bounded exploitability and lower-bounded reward against an opponent. Our method can exploit the weakness of its opponent without significantly sacrificing its exploitability. Empirical results show that our method significantly outperforms NE baselines when opponents play non-NE strategies and keeps low exploitability at the same time.
PDF Abstract