Advances in Predictive Control and Reinforcement Learning

The field of predictive control and reinforcement learning is moving towards more innovative and advanced approaches. Recent developments focus on improving the explainability and reliability of data-driven predictive control methods, as well as enhancing the safety and robustness of reinforcement learning algorithms. Notable advancements include the integration of bio-inspired reflexes into safe reinforcement learning methods and the use of contextual sampling to improve computational efficiency in data-enabled predictive control. Furthermore, researchers are exploring the application of Gaussian process optimization and system level synthesis to discover Koopman operators and develop more effective control policies.

Noteworthy papers include:

  • Towards explainable data-driven predictive control with regularizations, which provides a deeper understanding of regularization mechanisms in data-driven predictive control.
  • Bresa: Bio-inspired Reflexive Safe Reinforcement Learning for Contact-Rich Robotic Tasks, which proposes a novel safe reinforcement learning method inspired by biological reflexes.
  • Reinforcement learning for efficient and robust multi-setpoint and multi-trajectory tracking in bioprocesses, which presents a reinforcement-learning framework for multi-setpoint and multi-trajectory tracking in bioprocesses.
  • Less is More: Contextual Sampling for Nonlinear Data-Enabled Predictive Control, which introduces a novel data selection strategy to handle nonlinearities in data-enabled predictive control.
  • Inverted Gaussian Process Optimization for Nonparametric Koopman Operator Discovery, which leverages Gaussian process regression to develop a probabilistic Koopman linear model.
  • MPCritic: A plug-and-play MPC architecture for reinforcement learning, which presents a machine learning-friendly architecture that interfaces seamlessly with MPC tools.
  • Learning with Imperfect Models: When Multi-step Prediction Mitigates Compounding Error, which provides a rigorous analysis of the trade-off between single-step and multi-step prediction in learning-based control.
  • System Level Synthesis for Affine Control Policies: Model Based and Data-Driven Settings, which introduces a novel closed-loop parameterization for time-varying affine control policies.

Sources

Towards explainable data-driven predictive control with regularizations

Bresa: Bio-inspired Reflexive Safe Reinforcement Learning for Contact-Rich Robotic Tasks

Reinforcement learning for efficient and robust multi-setpoint and multi-trajectory tracking in bioprocesses

Less is More: Contextual Sampling for Nonlinear Data-Enabled Predictive Control

Inverted Gaussian Process Optimization for Nonparametric Koopman Operator Discovery

Preconditioned Additive Gaussian Processes with Fourier Acceleration

Semi-Data-Driven Model Predictive Control: A Physics-Informed Data-Driven Control Approach

Reinforcement learning for robust dynamic metabolic control

MPCritic: A plug-and-play MPC architecture for reinforcement learning

Learning with Imperfect Models: When Multi-step Prediction Mitigates Compounding Error

System Level Synthesis for Affine Control Policies: Model Based and Data-Driven Settings

Built with on top of