Towards Zero-Shot Autonomous Inter-Task Mapping through Object-Oriented Task Description

The successful use of Reinforcement Learning in complex tasks depends on techniques to scale-up classical learning algorithms because they suffer from the curse of dimensionality. Transfer Learning approaches have been used to accelerate learning by reusing knowledge gathered from the solution of previous tasks. However, discovering how different tasks are related is a very complex undertaking if a human is not available (or is unable) to manually establish a mapping between tasks. We here propose an algorithm to autonomously estimate a Probabilistic Inter-TAsk Mapping (PITAM) across tasks described in an object-oriented manner, which requires less domain knowledge than a handcrafted Inter-Task Mapping. We also propose two strategies for Temporal-Difference algorithms to transfer knowledge using learned PITAMs. Our experiments evaluate varied scenarios in which the source and target tasks differ in several aspects, and our proposal presents benefits over both regular learning and Q-value Reuse using a detailed InterTask Mapping.