How Knowledge about the Environment Influences the Performance of an Artificial Agent in a Multi-Agent Situation when trained using Reinforcement Learning
No Thumbnail Available
Robots are becoming increasingly popular in society, which increases the importance of them being complex cooperative and competitive agents. This study investigates whether knowledge about the environment improves the performance of an artificial agent in a multi-agent situation when trained using reinforcement learning. The influence of several types of knowledge has been researched using a simulation created in Unity. The game Capture the Flag poses as multi-agent situation, as it consists of multiple individual players whose objective it is to capture the other team's flag and bring it back safely to their own homebase. The agents are only rewarded for winning (and punished for losing), and have been trained with the use of several training strategies such as curriculum learning. Analysis of the results shows that knowledge has an influence on the winning rate of these artificial agents. The best performance was seen in teams of agents with access to knowledge about which team possesses which flags in combination with an extended visual range, though not all information turned out to be equally valuable. Still, all of the teams with added knowledge outperformed the teams without any added knowledge, and in most of them several distinct cooperative and competitive strategies have automatically emerged. This concludes that more knowledge improves the performance of an artificial agent in a multi-agent situation when trained using a reinforcement learning algorithm.
Faculteit der Sociale Wetenschappen