Safely Using Predictions in General-Sum Normal Form Games

It is often useful to predict opponent behavior when playing a generalsum two-player normal form game. However best-responding to an inaccurate prediction can lead to a strategy which is vulnerable to exploitation. This paper proposes a novel method, Restricted Stackelberg Response with Safety (RSRS), for an agent to select a strategy to respond to a prediction. The agent uses the confidence it has in the prediction and a safety margin which reflects the level of risk it is willing to tolerate to make a controlled trade-off between best-responding to the prediction and providing a guarantee of worst-case performance. We describe an algorithm which selects parameter values for RSRS to produce strategies that play well against the prediction, respond to a best-responding opponent, and guard against worst-case outcomes. We report results obtained by the algorithm on multiple general-sum games against different opponents.