2024/07/20 by Nanne A. Dieleman, Dieleman, Nanne A., Joost Berkhout +3
Computer Science · Mathematics · #FOS: Mathematics #Gaussian Processes and Bayesian Inference #Markov Chains and Monte Carlo Methods #Optimization and Control (math.OC) #Reinforcement Learning in Robotics
paper · pdf · doi:10.48550/arxiv.2407.14786
openalex publication_date 2024/07/20 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We develop a first-order (pseudo-)gradient approach for optimizing functions over the stationary distribution of discrete-time Markov chains (DTMC). We give insights into why solving this optimization problem is challenging and show how transformations can be used to circumvent the hard constraints inherent in the optimization problem. The optimization framework is model-free since no explicit model of the interdependence of the row elements of the Markov chain transition matrix is required. Upon the transformation we build an extension of Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm, called stochastic matrix SPSA (SM-SPSA) to solve the optimization problem. The performance of the SM-SPSA gradient search is compared with a benchmark commercial solver. Numerical examples show that SM-SPSA scales better which makes it the preferred solution method for large problem instances. We also apply the algorithm to the maximization of web-page rankings in web-graphs based on a real-life data set. As we explain in the paper, when applying a first-order gradient search one typically encounters a phenomenon which we call ``infliction points," that is, jumps in the optimization trajectories between periods of almost stationary behavior that slow down the optimization. We propose a heuristic for avoiding such infliction points and present a metastudy on a wide range of networks showing the positive effect of our heuristic on the convergence properties of SM-SPSA gradient search.