[A-AS02] From weather predictability to controllability

[AAS02-03] Controlling chaos through intermittent chaos enhancement

*Lin Li^1,2, Takemasa Miyoshi^1,2 (1.Prediction Science Laboratory, RIKEN Cluster for Pioneering Research, RIKEN, Japan, 2.Data Assimilation Research Team, RIKEN Center for Computational Science, Kobe, Japan.)

Keywords:controllability, chaotic systems

The sensitivity to initial conditions in chaotic dynamical systems makes them unpredictable, yet the same sensitivity implies highly efficient controllability. Such controls have been utilized in e.g., stabilizing turbulent convections, cardiac fibrillations, and chemical reactions. However, existing methods such as the OGY method rely on bi-directional perturbations of system parameters and may not be applied to many practical control tasks where parameters are only allowed to be perturbed uni-directionally. Here in this study, we train an intelligent controller through reinforcement learning, by which we can control the chaotic Lorenz-63 system by temporarily increasing the level of chaos. This is a counter-intuitive task that has not been attacked before. We then open the black box of the intelligent controller and undercover its control principle — unlike previous OGY methods that stabilize existing unstable periodic orbits, this new controller creates new stabilized periodic orbits. This study sheds light on chaos control tasks where the perturbation is constrained to a single direction.