Optimizing parameters in swarm intelligence using reinforcement learning: An application of Proximal Policy Optimization to the iSOMA algorithm

This paper presents a new algorithm for optimizing parameters in swarm algorithm using reinforcement learning. The algorithm, called iSOMA-RL, is based on the iSOMA algorithm, a population-based optimization algorithm that mimics the competition-cooperation behavior of creatures to find the optimal solution. By using reinforcement learning, iSOMA-RL can dynamically and continuously optimize parameters, which can play a crucial role in determining the performance of the algorithm but are often difficult to determine. The reinforcement learning technique used is the state-of-the-art Proximal Policy Optimization (PPO), which has been successful in many areas. The algorithm was compared to the original iSOMA algorithm and other algorithms from the SOMA family, showing better performance with only constant increase in computational complexity depending on number of function evaluations. Also we examine different sets of parameters to optimize and different reward functions. We also did comparison to widely used and state-of-the-art algorithms to illustrate improvement in performance over the original iSOMA algorithm.