Optimal Greedy Control in Reinforcement Learning.

Area of Science:

• Control Theory
• Robotics
• Computational Dynamics

Background:

• Optimal control problems often involve high-dimensional state spaces, posing computational challenges.
• Variational approaches and reinforcement learning are powerful tools for solving optimal control problems.
• Differential algebraic equations (DAEs) with nonlinearities and constraints are common in complex systems.

Purpose of the Study:

• To develop a method for dimensionality reduction in the state space of optimal control problems.
• To enhance the efficiency of reinforcement learning methods applied to complex control systems.
• To address control problems described by nonlinear differential algebraic equations.

Main Methods:

• The proposed method involves modifying Lagrange multipliers within a subset based on those from another subset.
• This technique is applied within the variational approach to optimal control.
• The method is implemented using the FRUND multibody system dynamics software package.

Main Results:

• Demonstrated successful dimensionality reduction for state space in optimal control.
• Validated the method's applicability through examples in robotics.
• Showcased effectiveness in vibration isolation for transport vehicles.

Conclusions:

• The proposed method offers an effective approach to state space dimensionality reduction in optimal control.
• This technique can improve the performance and applicability of reinforcement learning in complex dynamic systems.
• The implementation in FRUND facilitates practical application in engineering domains like robotics and vehicle dynamics.