Steering No-Regret Learners to Optimal Equilibria

We consider the problem of steering no-regret-learning agents to play desirable equilibria in extensive-form games via nonnegative payments. We show that steering is impossible if the total budget (across iterations) is finite. However, with average, realized payments converging to zero, we show that steering is possible. In the full-feedback setting, that is, when players' full strategies are observed at each timestep, it is possible with constant per-iteration payments. In the bandit-feedback setting, that is, when only trajectories through the game tree are observable, steering is impossible with constant per-iteration payments but possible if we allow the maximum per-iteration payment to grow with time, while maintaining the property that average, realized payments vanish. We supplement our theoretical positive results with experiments highlighting the efficacy of steering in large, extensive-form games, and show how our framework relates to optimal mechanism design and information design.