Adaptive optics control using model-based reinforcement learning

Area of Science:

• Astronomy
• Control Systems Engineering
• Artificial Intelligence

Background:

• Adaptive optics (AO) systems are crucial for high-resolution astronomical observations.
• Traditional AO control methods face challenges with temporal delays and calibration inaccuracies.
• Reinforcement learning (RL) presents a potential solution to these limitations.

Purpose of the Study:

• To investigate the application of model-based reinforcement learning (MBRL) for controlling AO systems.
• To evaluate the effectiveness of MBRL in addressing temporal delays and calibration errors in AO.
• To demonstrate MBRL's capability for continuous learning and adaptation in AO control.

Main Methods:

• Formulating the AO control loop as a model-based reinforcement learning (MBRL) problem.
• Implementing and simulating the MBRL approach on a Shack-Hartmann Sensor (SHS) based AO system.
• Utilizing a simulated AO system with 24 resolution elements.

Main Results:

• MBRL-controlled AO successfully predicted the temporal evolution of atmospheric turbulence.
• The MBRL system demonstrated an ability to adjust to mis-registration errors between the deformable mirror and SHS.
• Continuous learning was observed on timescales of seconds, enabling adaptation to changing conditions.

Conclusions:

• Model-based reinforcement learning is a promising approach for advanced AO system control in astronomy.
• MBRL can autonomously handle complex AO challenges like temporal delays and calibration issues.
• The adaptive nature of MBRL allows for robust performance under dynamic environmental conditions.