Symmetry reduction for deep reinforcement learning active control of chaotic spatiotemporal dynamics

Deep reinforcement learning (RL) is a data-driven, model-free method capable of discovering complex control strategies for macroscopic objectives in high-dimensional systems, making its application towards flow promising. Many systems interest possess symmetries that, when neglected, can significantly inhibit the and performance naive deep RL approach. Using test-bed consisting Kuramoto-Sivashinsky Equation (KSE), equally spaced actuators, goal minimizing dissipation power cost, we demonstrate that by moving problem to symmetry-reduced space, alleviate limitations inherent RL. We yields improved data efficiency as well policy efficacy compared policies found Interestingly, learned symmetry aware agent drives system toward an equilibrium state forced KSE connected continuation unforced KSE, despite having been given no explicit information regarding existence. I.e., achieve goal, algorithm discovers stabilizes system. Finally, robust observation actuation signal noise, parameters it has not observed before.