High-dimensional continuous action space control via trust region optimized deep reinforcement learning.

Area of Science:

• Deep Reinforcement Learning
• Adaptive Control Systems
• Robotics and Automation

Background:

• High-dimensional continuous action spaces pose significant challenges for traditional control methods.
• Existing deep reinforcement learning algorithms often struggle with sampling efficiency and real-time performance in complex industrial applications.

Purpose of the Study:

• To introduce a novel deep reinforcement learning framework, ATRPO-ACS, for adaptive control in high-dimensional continuous action spaces.
• To improve sampling efficiency, real-time performance, and reduce errors in trajectory tracking and constraint violations.

Main Methods:

• Developed the Adaptive Trust Region Policy Optimization for Action Space Compression (ATRPO-ACS) framework.
• Integrated distributed KL constraint optimization, manifold projection, and residual compensation.
• Utilized trust region strategies for policy optimization.

Main Results:

• Achieved significant improvements in sampling efficiency and real-time performance.
• Reduced trajectory tracking errors to within ±0.08 mm for robotic arms.
• Decreased microgrid scheduling costs by 28.5% and shortened production cycles in automotive welding.

Conclusions:

• ATRPO-ACS demonstrates superior performance in adaptive control tasks.
• The framework offers robust theoretical and technical support for real-time optimization in industrial intelligent systems.
• The approach effectively addresses challenges in high-dimensional continuous action spaces.