Online collaborative multi-agent reinforcement learning by transfer of abstract trajectories
Maarten van Someren, Mike Pool, Sanne Korzec · UvA-DARE (University of Amsterdam) · 2008
In this paper we propose a method for multi-agent reinforcement learning by automatic discovery of abstract trajectories. Local details are abstracted from successful trajectories and the resulting generalized, abstract trajectories are exchanged between agents. Each agent learns a policy for its own environment. By abstracting trajectories and sharing the result the agents benefit from each others learning. This reduces the overall learning time compared to individual learning.