STEER: When and How to Exploit Population-Specific Behavior in GLEE

Game-theoretic policies provide principled starting points for economic interaction, but the population encountered in specific environments such as GLEE may depart systematically from the populations implicit in theoretical models. These departures motivate exploiting alternative policies that are specialized for the encountered populations, as long as the risk of deploying them unconditionally in the face of weak, transient play patterns is mitigated. We present Strategic Theory and Evidence-gated Empirical Response (STEER), an agent that implements an evidence-gated strategic population model. For each configuration of GLEE bargaining, negotiation, or persuasion games, STEER pairs a population-independent THEORY policy with a configuration-specific EXPLOIT policy that is informed by a statistical population model (SPM) trained on historical GLEE games. It uses an anytime-valid evidence process to determine when a frozen EXPLOIT version can safely be deployed; in parallel, it periodically retrains EXPLOIT. We compare play with THEORY, always-on EXPLOIT, ungated as well as gated STEER in historical and live GLEE evaluations. Evidence-gated STEER achieves a large improvement in rating and a mean payoff improvement over alternatives that forgo population exploitation or evidence-gating, whose contributing factors we interpret through behavioral analysis.