Model reference adaptive search

In Chap. 4, we consider a global optimization approach, called model reference adaptive search (MRAS), which provides a broad framework for updating a probability distribution over the solution space in a way that ensures convergence to an optimal solution. After introducing the theory and convergence results in a general optimization problem setting, we apply the MRAS approach to various MDP settings. For the finite- and infinite-horizon settings, we show how the approach can be used to perform optimization in policy space. In the setting of Chap. 3, we show how MRAS can be incorporated to further improve the exploration step in the evolutionary algorithms presented there. Moreover, for the finite-horizon setting with both large state and action spaces, we combine the approaches of Chaps. 2 and 4 and propose a method for sampling the state and action spaces. Finally, we present a stochastic approximation framework for studying a class of simulation- and sampling-based optimization algorithms. We illustrate the framework through an algorithm instantiation called model-based annealing random search (MARS) and discuss its application to finite-horizon MDPs.