Abstract
Stackelberg Equilibria arise naturally in a range of popular learning problems, such as in security games or automated mechanism design, and have received increasing attention in the reinforcement learning literature recently. We present a general framework for implementing Stackelberg Equilibria search as a multi-agent RL problem, allowing a wide range of design choices. We discuss how previous approaches can be seen as specific instantiations of this framework. As a key insight, we note that the design space allows for approaches not previously seen in the literature, for instance by leveraging multitask and meta-RL techniques for follower convergence. We evaluate examples of novel approaches predicted by our framework experimentally on standard benchmark domains. Finally, we discuss directions for future work implied by our work.
Abstract (translated)
URL
https://arxiv.org/abs/2210.11942