The Online Discovery Problem and Its Application to Lifelong Reinforcement Learning

We study lifelong reinforcement learning where the agent extracts knowledge from solving a sequence of tasks to speed learning in future ones. We first formulate and study a related online discovery problem, which can be of independent interest, and propose an optimal algorithm with matching upper and lower bounds. These results are then applied to create a robust, continuous lifelong reinforcement learning algorithm with formal learning guarantees, applicable to a much wider scenarios, as verified in simulations.