Related Experiment Videos
Learning to control a complex multistable system.
1Department of Mathematics, Colorado State University, Weber Building, Fort Collins, Colorado 80523, USA. sabino@math.colostate.edu
Summary
This study shows how to control a complex mechanical rotor system with multiple stable states, even with significant noise. Reinforcement learning stabilizes the system at a desired state efficiently without needing system equations.
Area of Science:
- Nonlinear dynamics
- Complex systems control
- Mechanical engineering
Background:
- The investigated system is a periodically kicked mechanical rotor without gravity.
- This system exhibits complex multistable behavior with numerous competing attracting states.
- Previous work highlighted the system's complex dynamics and multistability.
Purpose of the Study:
- To investigate the control of a periodically kicked mechanical rotor in the presence of noise.
- To demonstrate the stabilization of the system at a desired attracting state despite high noise levels.
- To apply a data-based control method for efficient system stabilization.
Main Methods:
- Utilized a control algorithm for chaotic systems.
- Applied reinforcement learning to find an optimal control policy.
- Employed a data-based approach, requiring no prior knowledge of system dynamics.
Main Results:
- Successfully stabilized the mechanical rotor system at a desired attracting state.
- Demonstrated effective control even under high noise conditions.
- Achieved system stabilization in a minimum number of iterations.
Conclusions:
- It is possible to control complex multistable systems like the kicked rotor, even with significant noise.
- Reinforcement learning provides an effective data-based strategy for stabilizing such systems.
- The developed method offers a robust approach for controlling systems without needing their governing equations.