Lin Yang (Princeton) - Learn Policy Optimally via Efficiently Utilizing Data

Learn Policy Optimally via Efficiently Utilizing Data

Recent years have witnessed increasing empirical successes in reinforcement learning. Nevertheless, it is an irony that many theoretical problems in this field are not well understood even in the most basic setting. For instance, the optimal sample and time complexities of policy learning in finite-state Markov decision process still remain unclear. Given a state-transition sampler, we develop a novel algorithm that learns an approximate-optimal policy in near-optimal time and using a minimal number of samples. The algorithm makes updates by processing samples in a “streaming” fashion, which requires small memory and naturally adapts to large-scale data. Our result resolves the long-standing open problem on the sample complexity of Markov decision process and provides new insights on how to use data efficiently in learning and optimization.

The algorithm and analysis can be extended to solve two-person stochastic games and feature-based Markov decision problems while achieving near-optimal sample complexity. We further illustrate several other examples of learning and optimization over streaming data, with applications in accelerating Astrophysical discoveries and improving network securities.