Mechanism Design as Inverse Games: A Computational Approach to Dynamic Mechanism Design

ICLR 2026 Conference Submission22095 Authors

Keywords: Dynamic Mechanism Design, Optimal Mechanism Design, Partially Observable Markov Games, Inverse Game Theory, Mechanism Design, Deep Reinforcement Learning

Abstract:

Mechanism design is often described as inverse game theory: rather than analyzing equilibria of a game, the designer specifies rules to induce desirable outcome at equilibrium. We present a computational framework for optimal dynamic mechanism design with evolving agent types. We cast the problem as a constrained optimization over partially observable Markov games, with incentive compatibility and individual rationality encoded as constraints. To solve it, we develop a min–max optimization approach and propose two methods for handling partial
observability: (i) Bayesian belief-state tracking with convergence guarantees in discrete-type settings, and (ii) recurrent neural embeddings that scale to continuous types. In bandit auction experiments, our framework recovers known single-item benchmarks and discovers new incentive-compatible mechanisms in multi-item environments lacking analytical solutions.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 22095