Thursday ML Seminar

Towards a Rigorous Approach to Approximate Computations in Huge-Scale Markovian Decision Processes

In this talk, without assuming any background beyond familiarity with basic probability and linear algebra, I will describe exciting new results on the computation of near-optimal policies in huge-scale Markovian decision processes (MDPs) via the so-called approximate linear programming methodology. In a nutshell, the new results provide meaningful upper bounds on the quality of the computed policies as a function of the linear basis functions chosen by the algorithm designer, while avoiding unrealistic assumptions which have effectively become the norm in prior results. As opposed to previous work that relied on worst-case reasoning over randomly sampled constraints, in this work the analysis is done using an operator-theoretic approach. I will finish by discussing the remaining main open problems, connections to alternative approaches, as well as potential extensions. Markov decision processes lie at the hear of many reinforcement learning algorithms and have found numerous applications across many engineering and scientific problems.
Based on joint work with Chandrashekar Lakshminarayanan and Shalabh Bhatnagar