Applications of Cellular Learning Automata and Reinforcement Learning in Global Optimization

Abstract

Various scientific and engineering problems can be modeled as optimization tasks. Roughly speaking, optimization problems can be categorized as static problems and dynamic problems. Among different optimization approaches, nature-inspired methods such as evolutionary algorithms have attracted considerable interest in recent years. A nature-inspired algorithm is an iterative process. Much effort has been put into the development of versatile nature-inspired optimization methods. However, it is well accepted that a general-purpose universal optimization algorithm is impossible: no strategy or scheme can outperform the others on all possible optimization problems. Besides, a single evolutionary scheme or operator may always follow similar trajectories. Therefore, it would be better to use diverse evolutionary operators to increase the chance of finding optima. Moreover, the parameter and strategy configuration of an optimization algorithm can significantly affect its performance. On a specific problem, proper configurations can result in high-quality solutions and high convergence speeds. Therefore, the utilization of multiple strategies or techniques, along with a suitable adaptation mechanism, can significantly enhance a nature-inspired optimization method. Indeed, some recently developed nature-inspired optimization methods incorporate a collection of strategies in some of their algorithmic elements such as mutation, crossover, or neighborhood structure. In this chapter, we will show how cellular learning automata and reinforcement learning can be utilized for the adaptive selection of proper strategies in nature-inspired optimization algorithms.