Kinetic theory of decentralized learning for smart active matter

Abstract: Smart active matter describes agents which can process information according to their individual policy to solve predefined tasks. We introduce a theoretical framework to study a decentralized learning process in which agents can locally exchange their policy to optimize their reward and thus attain target macroscopic states. We use our formalism to derive explicit hydrodynamic equations for the policy dynamics. We apply the theory to two different microscopic models where policies correspond either to fixed parameters similar to evolutionary dynamics, or to state-dependent controllers known from the field of robotics. We find good agreement between theoretical predictions and agent-based simulations. By deriving fundamental control parameters and uncertainty relations, our work lays the foundations for a statistical physics analysis of decentralized learning.