Imitation learning-driven approximation of stochastic control models

Predictive control is a widely adopted methodology in numerous industries to manage multi-variable control problems under constraints. Traditional predictive control methods, however, assume complete model knowledge and deterministic behavior, which is often unrealistic in practical applications due to uncertainties and disturbances. Stochastic predictive control addresses these limitations by incorporating probabilistic models to enhance robustness in uncertain environments, but this approach comes at the cost of significantly increased computational complexity, making real-time implementation challenging. This paper proposes a method based on imitation learning to approximate the solution of stochastic predictive control, significantly reducing computational burden while maintaining predictive capabilities and robustness. The effectiveness of the proposed method is first demonstrated on the cart-pole stabilization problem, followed by its application to optimal lithium-ion battery charging. Results from both case studies underscore the method's robust approximation capabilities and substantial computational cost reduction, underscoring its potential for real-time applications in diverse domains, including robotics, energy management, and autonomous systems.