10th International Congress on Information and Communication Technology

Authors - Iaroslav Omelianenko
Abstract - Neuroevolution algorithms need to evaluate at the end of each epoch the fitness scores of each organism in a population of solvers within the problem space where a solution is sought. This evaluation often involves running complex environmental simulations, which can significantly slow down the training speed if done sequentially. This work proposes a solution that utilizes the inherent capabilities of the Go programming language to run complex simulations in local parallel processes (routines). The efficiency of this proposed solution is compared to sequential evaluation using two classic reinforcement learning experiments, specifically single and double pole balancing. Direct comparisons indicate that the proposed solution is up to five times faster than the sequential approach when complex environmental simulations are required for objective function evaluation.