Evolutionary reinforcement learning of dynamical large deviations.

Area of Science:

• Computational Physics
• Machine Learning
• Statistical Mechanics

Background:

• Dynamical large deviations are crucial for understanding rare events in stochastic systems.
• Calculating these deviations often involves computationally intensive methods.
• Existing frameworks may not fully capture the complexities of path-extensive quantities.

Purpose of the Study:

• To develop a novel method for bounding and calculating the likelihood of dynamical large deviations.
• To leverage evolutionary reinforcement learning for analyzing stochastic models.
• To bridge the gap between physics problems and machine learning frameworks.

Main Methods:

• An agent, representing a stochastic model, propagates continuous-time Monte Carlo trajectories.
• Rewards are assigned based on the values of path-extensive quantities.
• Evolutionary algorithms optimize agents to improve the calculation of large-deviation rate functions.
• For large state spaces, neural networks parameterize the model's rates.

Main Results:

• Demonstrated the feasibility of using evolutionary reinforcement learning to bound and calculate dynamical large deviations.
• Showcased the method's applicability to models with varying state space sizes.
• Successfully linked path-extensive physics problems to a machine learning framework.

Conclusions:

• Evolutionary reinforcement learning offers a powerful new approach for tackling complex problems in statistical mechanics and physics.
• This framework facilitates the computation of large-deviation rate functions, previously a significant challenge.
• The study highlights the potential of integrating advanced machine learning techniques into physical modeling.