Data-Based Predictive Control via Multistep Policy Gradient Reinforcement Learning.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Model-free predictive control (MFPC) offers an alternative to traditional methods by avoiding explicit system models.
• Data-driven approaches are increasingly important for real-time systems.
• Reinforcement learning (RL) provides a powerful framework for learning optimal control policies.

Purpose of the Study:

• To present a novel model-free predictive control algorithm for real-time systems.
• To leverage multistep policy gradient reinforcement learning for performance improvement.
• To develop a data-driven control strategy that does not require prior knowledge of system dynamics.

Main Methods:

• The algorithm utilizes multistep policy gradient reinforcement learning.
• Cooperative games model predictive control as multiagent optimization problems.
• Neural networks approximate the action-state value function and control policy.
• Weighted residual methods determine network weights.

Main Results:

• The proposed algorithm effectively improves real-time system performance.
• Model-free design is achieved by learning from offline and real-time data.
• Optimality of the predictive control policy is guaranteed through the game-theoretic approach.

Conclusions:

• The developed model-free predictive control algorithm is effective for real-time systems.
• The data-driven, reinforcement learning-based approach eliminates the need for system dynamics knowledge.
• The integration of neural networks and cooperative games ensures efficient and optimal control.