Lifelong Learning-Based Optimal Trajectory Tracking Control of Constrained Nonlinear Affine Systems Using Deep Neural...

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Machine Learning

Background:

• Optimal trajectory tracking is crucial for robotic systems.
• Uncertain nonlinear systems with state constraints pose significant control challenges.
• Lifelong learning is essential for adapting control policies in dynamic environments.

Purpose of the Study:

• To develop a novel lifelong integral reinforcement learning (LIRL)-based optimal trajectory tracking scheme.
• To address challenges in uncertain nonlinear continuous-time (CT) affine systems with state constraints.
• To improve control policy generation and mitigate catastrophic forgetting in multitasking systems.

Main Methods:

• Utilized a critic multilayer neural network (MNN) or Deep NN to approximate value functions and generate optimal control policies.
• Employed a singular value decomposition (SVD)-based method for online tuning of critic MNN weights.
• Incorporated an online lifelong learning (LL) scheme to prevent catastrophic forgetting.
• Addressed state constraints using a time-varying barrier function (TVBF).

Main Results:

• Achieved optimal trajectory tracking for uncertain nonlinear CT affine systems.
• Demonstrated effective mitigation of catastrophic forgetting in multitasking systems.
• Successfully handled state constraints through the TVBF.
• Showcased uniform ultimate boundedness (UUB) of the closed-loop system via Lyapunov stability analysis.

Conclusions:

• The proposed LIRL-based optimal framework effectively addresses trajectory tracking in constrained nonlinear systems.
• The novel SVD-based tuning and LL scheme enhance control policy adaptability and robustness.
• Experimental results on a two-link robotic manipulator show a 47% total cost reduction, validating the method's practical effectiveness.