Development of General-purpose Optimization Methods Based on Bandit Algorithm

Abstract

The goal of this study was to develop the efficient optimization methods with roughly estimating the values in the areas and considering the exploration and exploitation in the optimization problem. First, we proposed Grid Area Search (GAS) algorithm as a base of our methods by upgrading the random search (RS) algorithm which is the simplest algorithm for the mathematical programming problem. Specifically, our methods convert the mathematical programming problem to the Multi-armed bandit problem by dividing search space. Therefore, it is assumed that the coordination of the exploration and exploitation can be introduced by the UCB algorithms. Then we also proposed UCB-GAS and UCT-GAS. We performed comparative experiments in continuous optimization problems to make sure its convergence speed.