Reinforcement learning enabled control of chaotic dynamics

We illustrate the utility of Deep Reinforcement Learning (RL) for controlling chaotic systems. A deep RL policy network, based on proximal policy optimization, is employed to stabilize the unstable fixed points and periodic orbits embedded in the chaotic-attractor of the Lorenz system. Previous attempts to control the underlying chaotic trajectories have relied on linearization of the dynamics around the targeted solutions, or on time-delayed feedback based on the output variables. However, there are certain caveats associated with these control approaches such as, requiring an a priori understanding of the chaotic system to be controlled, and difficulty in stabilizing certain classes of periodic orbits. These issues are overcome in the RL enabled control, especially when long term correlations are accounted for using layers of Long Short Term Memory (LSTM) cells in the policy network.