Online Inverse Optimal Control for Time-Varying Cost Weights

Area of Science:

• Control Theory
• Machine Learning
• Robotics

Background:

• Traditional inverse optimal control assumes constant cost weights, limiting real-world applicability.
• Many practical scenarios involve dynamic cost functions that change over time.
• Recovering these time-varying cost weights is a significant challenge in control systems.

Purpose of the Study:

• To develop an adaptive online inverse optimal control method for recovering time-varying cost weights.
• To address the limitations of existing methods that assume constant cost parameters.
• To enable more accurate modeling of expert behavior in dynamic environments.

Main Methods:

• Utilizing neural network approximation for modeling time-varying cost weights.
• Conducting a well-posedness analysis to ensure problem solvability.
• Proposing a novel weight updating law for stable convergence of solutions.
• Validating the approach through simulations on a linear system.

Main Results:

• Demonstrated the effectiveness of the proposed adaptive online inverse optimal control strategy.
• Established conditions for the uniqueness of neural network weights for the inverse optimal control problem.
• Ensured stability and convergence of the solutions through a proposed updating law.
• Successfully recovered time-varying cost weights in simulation.

Conclusions:

• The proposed neural network-based adaptive online inverse optimal control effectively handles time-varying cost weights.
• The method offers a robust solution for real-time inverse optimal control problems.
• Applicable to a broad range of dynamic systems and expert demonstrations.
• Advances the field of inverse optimal control for complex, real-world applications.