Continual Multi-agent Reinforcement Learning applied in cooperative games

Reinforcement Learning combined with Continual Learning offers the poten- tiality to explore many of the challenges of the modern era in almost every field of the society, from financial and portfolio management to human-robot coop- eration in the industry 4.0 but also in the healthcare and in smart grids and energy management. The intersection of reinforcement learning and continual learning in the multi-agent scenario has not been extensively investigated. It is therefore valuable to explore this crossover. In particular, this thesis analyzes the application of continual learning techniques to enhance the adaptability of multi-agent reinforcement learning systems in dynamic and cooperative envi- ronments. Among the issues that continual learning aims at overcoming stands catastrophic forgetting, namely the fact that agents lose previously acquired knowledge when learning new tasks or adapting to changing scenarios. This research aims to understand if it is possible to overcome this limitation, in par- ticular in the scenario of fully cooperative games, by implementing and testing some architectures that allow agents to seamlessly transition between differ- ent teams while maintaining the ability to effectively perform in previously encountered ones.