Lipschitz Lifelong Reinforcement Learning

We consider the problem of knowledge transfer when an agent is facing a series Reinforcement Learning (RL) tasks. introduce novel metric between Markov Decision Processes and establish that close MDPs have optimal value functions. Formally, functions are Lipschitz continuous with respect to tasks space. These theoretical results lead us value-transfer method for Lifelong RL, which we use build PAC-MDP algorithm improved convergence rate. Further, show experience no negative high probability. illustrate benefits in RL experiments.