An Efficient Dynamic Optimization Algorithm for Path-Constrained Switched Systems.

Area of Science:

• Control Systems Engineering
• Optimization Theory
• Reinforcement Learning

Background:

• Model-based adaptive reinforcement learning is crucial for industrial systems with switching mechanisms.
• Existing methods often involve complex multi-stage algorithms with significant computational burden.
• Ensuring path constraints during optimization is a key challenge for switched systems.

Purpose of the Study:

• To develop an efficient dynamic optimization strategy for switched systems.
• To guarantee satisfaction of path constraints throughout the entire time period.
• To reduce the number of nonlinear programs (NLPs) and computational burden.

Main Methods:

• A single-level algorithm is proposed, evaluating objective function gradients using adjoint systems and sensitivity equations.
• Optimization of system input and switch times is performed concurrently within each iteration.
• A novel policy iteration method adapted for switched systems ensures solution feasibility.

Main Results:

• The algorithm significantly reduces the number of NLPs and computational burden compared to multistage algorithms.
• Feasibility of the optimal solution is guaranteed.
• The proposed algorithm demonstrates finite termination and convergence to a Karush-Kuhn-Tucker (KKT) condition-satisfying solution.
• Numerical studies confirm less computational time compared to bi-level algorithms.

Conclusions:

• The presented single-level dynamic optimization strategy is efficient and computationally less expensive for switched systems.
• The method guarantees path constraint satisfaction and converges to feasible solutions.
• This approach offers a practical advancement for optimizing industrial systems with switching mechanisms.