Abstraction & Meta-Reinforcement Learning

Reinforcement learning is hard in a fundamental sense: even in finite and deterministic environments, it can take a large number of samples to find a near-optimal policy. In this talk, I discuss the role that abstraction can play in achieving reliable yet efficient learning and planning. I first introduce classes of state abstraction that induce a trade-off between optimality and the size of an agent’s resulting abstract model, yielding a practical algorithm for learning useful and compact representations from a demonstrator. Moreover, I show how these learned, simple representations can underlie efficient learning in complex environments. Second, I analyze the problem of searching for options that make planning more efficient. I present new computational complexity results that illustrate it is NP-hard to find the optimal options that minimize planning time, but show this set can be approximated in polynomial time. Collectively, these results provide a partial path toward abstractions that minimize the difficulty of high quality learning and decision making.