Transfer Reinforcement Learning Across Tasks with Dynamics Mismatch

Reinforcement Learning has recently emerged as a viable solution for various sequential decision-making problems. However, standard Reinforcement learning agents require a large number of samples to learn effectively and transfer learning is a mechanism aimed at alleviating this issue. When transferring knowledge from already mastered source tasks to a new target task, the similarity between the source and target task can play a crucial role in determining the usefulness of transfer. This thesis showcases the utility of one such measure of task similarity, based on the upper bound of the differences in optimal state-action values, as a gatekeeper for selective transfer in the transfer learning setting where the previously solved tasks and current task differ with respect to the transition dynamics. Numerical results, including a novel application of transfer learning for epidemic spread control, demonstrate the benefits of preventing negative transfer while accelerating learning otherwise.M.A.S