Nonconvergence to Saddle Boundary Points under Perturbed Reinforcement Learning
Abstract
This paper presents a novel reinforcement learning algorithm and provides conditions for global convergence to Nash equilibria. For several classes of reinforcement learning schemes, including the ones proposed here, excluding convergence to action profiles which are not Nash equilibria may not be trivial, unless the step-size sequence is appropriately tailored to the specifics of the game. In this paper we sidestep these issues by introducing a perturbed reinforcement learning scheme where the strategy of each agent is perturbed by a strategy-dependent perturbation (or mutations) function. Contrary to prior work on equilibrium selection in games where perturbation functions are globally state dependent, the perturbation function here is assumed to be local, i.e., it only depends on the strategy of each agent. We provide conditions under which the strategies of the agents will converge to an arbitrarily small neighborhood of the set of Nash equilibria almost surely. This extends prior analysis on reinforcement learning in games which has been primarily focused on urn processes. We finally specialize the results to a class of potential games.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 07, 2012
- Accession Number
- ADA584441
Entities
People
- Anders Rantzer
- Georgios C. Chasparis
- Jeff S. Shamma
Organizations
- Georgia Tech