Department of Industrial Engineering & Decision Analytics [Joint IEDA/ISOM] seminar - Multi-agent Adaptive Mechanism Design

We study a sequential mechanism design problem where a principal elicits truthful reports from multiple rational agents without prior knowledge of their beliefs. We propose Distributionally Robust Adaptive Mechanism, which integrates mechanism design and online learning to jointly ensure truthfulness and cost efficiency. The mechanism iteratively estimates agents’ beliefs and solves a distributionally robust program with shrinking ambiguity sets. It achieves high-probability truthfulness and low regret, with a matching lower bound. The algorithm can be generalized to any plug-in estimators supporting structured priors and delayed feedback. To our knowledge, this is the first adaptive mechanism under the general settings that maintains truthfulness and achieves optimal regret when incentive constraints are unknown and must be learned.