Multi-agent reinforcement learning is a promising approach for optimizing the behavior of multiple agents with minimal expert guidance. Such agents can be, for example, co-operating robots or wireless devices. The goal of the project is to increase understanding on how to control multi-agent learning. The learning should be fast but not sacrifice the quality of the end solution. For making multi-agent reinforcement learning more efficient and principled this project focuses on: (i) Avoiding expensive data collection in the operating environment by utilizing computational models to predict future events. (ii) Developing new methods with the aim to increase computational efficiency. The new methods start from easy tasks and progress to the actual hard task automatically. (iii) Planning to collect valuable data for both model learning and improved quality of agent behavior.

2023

2027