Abstract:
A common problem in Reinforcement Learning (RL) is that the reward function is hard to express. This can be overcome by resorting to Inverse Reinforcement Learning (IRL), which consists in first obtaining a reward function from a set of execution traces generated by the expert agent, and then making the learning agent learn the expert's behavior --this is known as Transfer Learning (TL). Typical IRL solutions rely on a numerical representation of the reward function, which raises problems related to the adopted optimization procedures.
We describe a TL method where the execution traces generated by the expert agent, possibly via planning, are used to produce a logical (as opposed to numerical) specification of the reward function, to be incorporated in a device known as Restraining Bolt (RB). The RB can be attached to the learning agent to drive the learning process and ultimately make it imitate the expert. We show that TL can be applied to heterogeneous agents, with the expert, the learner and the RB using different representations of the environment's actions and states, without specifying mappings among their representations.
