Incorporating Prior Knowledge and Previously Learned Information into Reinforcement Learning Agents
Kevin R. Dixon, Richard Malak, Pradeep K. Khosla · 2000
Reinforcement learning has received much attention in the past decade. The primary thrust of this research has focused on tabula rasa learning methods. That is, the learning agent is initially unaware of its environment and must learn or re-learn everything. We feel that this is neither realistic nor effective. While the agent may start out with little or no knowledge of its environment, it must be able to incorporate new information into the learning of subsequent tasks otherwise the learning effort is largely wasted. To address the shortcomings of tabula rasa learning, we present a general and intuitive approach for incorporating previously learned information and prior knowledge into the reinforcement learning process. We demonstrate the potential of this method on learning problems in the mobilerobot and grid-world domains, where results indicate that learning time can be decreased. We also demonstrate that multiple knowledge sources can be incorporated into the learning process. 1