On February 9th we were notified that the following project is granted by NWO (open competition exact sciences)

Analyzing and controlling Multi-Agent Learning (AniMAL)

Multi-Agent Systems (MAS) are accepted to be an important method for solving problems of a distributed nature. Key to the success of MAS is efficient and effective Multi-Agent Learning (MAL). In this project we address two important research challenges in MAL. 1) Current learning theory for single agents does not extend to MAL. Convergence guarantees no longer hold and there exists no general formal theory describing and elucidating the conditions under which algorithms for MAL are successful. 2) It is currently an open question how to handle the dynamics of MAL for many states, many agents and continuous strategy spaces. We address both challenges using an innovative combination of Evolutionary Game Theory (EGT) and Reinforcement Learning. More precisely, we will provide a theoretical backbone for MAL based on the replicator equations of EGT. In addition to understanding MAL, this theory will allow for fast evaluation of the success of parameters and algorithms under specific learning conditions. Furthermore we will extend this theoretical framework, such that it can deal with continuous strategy spaces, which avoids cumbersome discretization processes.