arXiv Board & Classic Game AI· Miaobo Hu·· 2026-10-15AI score28
SafeAdapt-Poker: Counterfactual risk calibration for budgeted adaptation under opponent shift
SafeAdapt-Poker: Counterfactual Risk Calibration for Budgeted Adaptation under Opponent Shift
AI summary
SafeAdapt-Poker proposes a budgeted adaptive control method for shifts in opponent behavior. It decides how far to deviate from a frozen anchor policy by weighing counterfactual policy risk against budget consumption. On exact Leduc poker, it achieves a matched weak-opponent-pool gain of +0.0718 and a worst-opponent difference of -0.0126. Under the same session budget, its delayed opportunity capture rate rises from 0.6318 under myopic allocation to 0.7841. In head-to-head matches against abstracted HUNL, it reaches +42.7 mbb/hand.
Source: arXiv Board & Classic Game AI · arxiv.org