Connections between reinforcement learning and statistical mechanics
Sample walks generated by the dynamics learnt in trying to efficiently sample random walker excursions.The aim of this project is to study the numerous and deep connections between reinforcement learning (RL) and statistical mechanics (SM). This two way relationship will guide the application of RL to problems in SM, in particular nonequilibrium SM, and allow an alternative perspective on RL itself through the lens of SM techniques.
Our first work in this line of research exploits a deep relationship between sampling probabilistically rare trajectories of dynamical processes and maximum entropy RL, allowing application of RL to the construction of an efficient sampling dynamics for rare trajectories in discrete time. This is now available on ArXiv.
Dominic C. Rose
Research Fellow in Theoretical Physics
My research interests include nonequilibrium physics, large deviations and reinforcement learning.
Publications
A reinforcement learning approach to rare trajectory sampling
