An Approximate Newton Method for Markov Decision Processes

TL;DR

This paper introduces a novel approximate Newton algorithm for optimizing Markov Decision Processes, offering improved convergence, robustness, and computational efficiency over traditional methods, with theoretical guarantees and practical advantages.

Contribution

The paper presents a new approximate Newton method with guaranteed negative-semidefinite Hessian, efficient inference, and sparsity properties, advancing optimization techniques for MDPs.

Findings

Algorithm exhibits excellent convergence properties

Method is robust to various problem settings

Approximate Hessian enables efficient computation

Abstract

Gradient-based algorithms are one of the methods of choice for the optimisation of Markov Decision Processes. In this article we will present a novel approximate Newton algorithm for the optimisation of such models. The algorithm has various desirable properties over the naive application of Newton's method. Firstly the approximate Hessian is guaranteed to be negative-semidefinite over the entire parameter space in the case where the controller is log-concave in the control parameters. Additionally the inference required for our approximate Newton method is often the same as that required for first order methods, such as steepest gradient ascent. The approximate Hessian also has many nice sparsity properties that are not present in the Hessian and that make its inversion efficient in many situations of interest. We also provide an analysis that highlights a relationship between our approximate Newton method and both Expectation Maximisation and natural gradient ascent. Empirical results suggest that the algorithm has excellent convergence and robustness properties.