AML Talk

The objective of this work is to improve the performance of a reinforcement learning (RL) agent through the transfer of prior knowledge from a repository. The repository may contain both previously learned and hand-coded information. Such an approach is desirable because it promotes reuse of prior learning experience and enables the inclusion of domain knowledge.

The literature has demonstrated proficiency for knowledge transfer within a limited setting. However, most approaches reported are bound to a particular representation of information. The results presented here differ in that a knowledge source is treated as a black box; only an interface to the knowledge is assumed.

An RL agent learns though trial-and-error interactions with its environment. This trial-and-error process, known as exploration, can be very time consuming - an agent may take many unproductive actions before attempting the correct one. The work to be presented uses prior knowledge to improve exploration efficiency. Qualatatively, the RL agent is biased toward actions thought by the prior knowledge to be useful and away from those thought to be unproductive.

The benefits and shortcomings of this approach will be demonstrated and discussed in the context of a competitive virtual robot learning task. Possible extensions of the work will be discussed.