Hierarchical Reinforcement Learning for Non-Stationary Environments
Rachel Haighton, Amirhossein Asgharnia, Howard M. Schwartz, Sidney Nascimento Givigi · 2023
What indications are there when the environment changes and the learned policy is no longer optimal? Is it possible to predict when a non-stationary environment changes in some way? In this paper we propose a method that helps agents know when to retrain their policies via reinforcement learning. The agents detect changes based on the temporal difference. A hierarchical learning model is used to aid in these non- stationary environments. The hierarchical learning model has two levels, the higher-level policy, which we call the learning switch, and the lower-level policy, which tells the agents their suitable action to play the game. The higher-level policy determines when reinforcement learning should be turned on or off based on the temporal differences calculated within a game or episode. Two multi-agent differential games are used as examples. The first two examples tackle the problem in cooperative games, while the last example addresses the competitive game. The results show that the agents can maintain suitable performance by switching on the learning process for a few iterations after environment changes occurs. In this paper we consider the change in environment to be long term occurrence within the dynamics of games; for example the mass of an agent may become heavier.