Bifurcation Analysis for Reinforcement Learning Agents
This project aims at using bifurcation analysis techniques for the study of the dynamics of reinforcement learning agents.
The application of reinforcement learning algorithms to multiagent domains may cause complex non-convergent dynamics. The replicator dynamics, commonly used in evolutionary game theory, proved to be effective for modeling the learning dynamics in normal form games. Nonetheless, it is often interesting to study the robustness of the learning dynamics when either learning or structural parameters are perturbed. This is equivalent to unfold the catalog of learning dynamical scenarios that arise for all possible parameter settings which, unfortunately, cannot be obtained through ``brute force simulation of the replicator dynamics. The analysis of bifurcations, i.e., critical parameter combinations at which the learning behavior undergoes radical changes, is mandatory. In this work, we introduce a one-parameter bifurcation analysis of the Selten's Horse game in which the learning process exhibits a set of complex dynamical scenarios even for relatively small perturbations on payoffs.
We are now investigating how to apply bifurcation analysis for the study of the dynamics of simple reinforcement learning sysmtes.
Selten's Horse Game
Changing the Replicator Dynamics
- Alessandro Lazaric, PhD Student
- Josè Enrique Munoz, PhD Student
- Fabio Dercole, PhD
- Marcello Restelli, PhD
A. Lazaric, E. Munoz de Cote, F. Dercole, M. Restelli - Bifurcation Analysis of Reinforcement Learning Agents
- Adaptive and Learning Agents and Multi-Agent Systems (ALAMAS) pp. 111--125, Maastricht, The Netherlands, April, 2007
- BibtexAuteur : A. Lazaric, E. Munoz de Cote, F. Dercole, M. Restelli
Titre : Bifurcation Analysis of Reinforcement Learning Agents
Dans : Adaptive and Learning Agents and Multi-Agent Systems (ALAMAS) -
Adresse : Maastricht, The Netherlands
Date : April 2007