Evolving population method for real-time reinforcement learning

Abstract

Reinforcement learning has recently been recognized as a promising means of machine learning, but its applicability remains limited in real-time environment due to its short response time, high computational complexity, and instability in learning. Although researchers devised several measures in attempts to press beyond the horizon, the problems consisting of large branching factors with real-time properties still stays unconquered, demanding a new method for reinforcement learning as a whole. In this paper, we propose Evolving Population. This method improves the performance of reinforcement learning by optimizing hyperparameters and available actions. This method uses an iterative structure based on an evolutionary strategy to optimize these elements. We validate the performance of our method in an environment with real-time properties and large branching factors.