Discrete-Time Reinforcement Learning Adaptive Control for Non-Gaussian Distribution of Sampling Intervals.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Unknown discrete-time systems present control challenges.
• Non-Gaussian sampling intervals complicate controller design.
• Reinforcement learning offers a potential solution for adaptive control.

Purpose of the Study:

• To develop an optimal reinforcement learning (RL) controller for unknown discrete-time systems.
• To address challenges posed by non-Gaussian sampling interval distributions.
• To enhance system performance in terms of tracking and compensation.

Main Methods:

• Utilized actor and critic networks with MiFRENa and MiFRENc architectures.
• Developed a learning algorithm with convergence analysis for learning rates.
• Employed estimated co-state for improved learning laws.
• Implemented weight transfer omission in the critic network for specific scenarios.

Main Results:

• The proposed RL controller demonstrated superior performance compared to a benchmark controller, especially for non-Gaussian distributions.
• The learning laws effectively improved dead-zone compensation and handling of nonlinear variations.
• Convergence analysis ensured stable learning rates based on internal signals and tracking errors.

Conclusions:

• The novel RL-based optimal controller is effective for unknown discrete-time systems with non-Gaussian sampling intervals.
• The proposed method offers significant improvements in system performance and robustness.
• The approach provides a viable solution for complex control problems in adaptive systems.