Caltech researchers show how to leverage imperfect AI advice in repeated games
New pseudo-metric quantifies advice usefulness, enabling efficient Stackelberg strategies against no-regret learners.
A team from Caltech (Handina, Li, Panaganti, Mazumdar, Wierman) published a paper at AISTATS 2026 analyzing how an agent can effectively use imperfect machine-learned advice in repeated strategic interactions against opponents who use no-regret learning algorithms. The researchers introduce a pseudo-metric to characterize the usefulness of advice instances, applying it to two common forms: simulators that can approximate opponents' strategies, and payoff matrix predictions. They prove that with correctness guarantees on the advice, a player can compute approximate Stackelberg strategies more efficiently—reducing the number of interactions traditionally needed.
When advice has no correctness guarantees, the team shows a fundamental trade-off: a player cannot simultaneously guarantee near-Stackelberg performance when advice is accurate and a no-regret condition when advice is inaccurate. However, they demonstrate that an advice-aided player can still achieve outcomes that weakly dominate their utility in some (coarse)-correlated equilibria. This work bridges machine learning advice and game theory, offering practical insights for AI agents operating in competitive environments where opponent behavior is learned over time.
- Introduced a novel pseudo-metric to quantify the usefulness of imperfect advice in repeated games against no-regret learners.
- With guaranteed-accurate advice (simulators or payoff predictions), players can compute approximate Stackelberg strategies with reduced interaction complexity.
- Without guarantees, near-Stackelberg and no-regret cannot be simultaneously ensured, but advice can yield weakly dominant outcomes in correlated equilibria.
Why It Matters
Provides a theoretical foundation for AI agents to use imperfect advice in competitive settings, enabling more efficient strategic decision-making.