A Systematic Review and Taxonomy of Reinforcement Learning-Model Predictive Control Integration for Linear Systems
A new systematic review reveals 5 key RL roles in MPC architectures for linear control.
A team of six researchers—Mohsen Jalaeian Farimani, Roya Khalili Amirabadi, Davoud Nikkhouy, Malihe Abdolbaghi, Mahshad Rastegarmoghaddam, and Shima Samadzadeh—has released a comprehensive systematic literature review on the integration of Reinforcement Learning (RL) and Model Predictive Control (MPC) for linear and linearized systems. Published on arXiv (2604.21030) on April 22, 2026, this paper covers peer-reviewed and indexed studies up to 2025, offering a much-needed taxonomy for a rapidly growing but fragmented field.
The review proposes a multi-dimensional taxonomy covering five key aspects: RL functional roles (e.g., direct policy optimization, adaptation, constraint enforcement), RL algorithm classes (e.g., Q-learning, policy gradients, actor-critic), MPC formulations (e.g., linear, quadratic, robust), cost-function structures, and application domains (e.g., robotics, process control, autonomous vehicles). Through cross-dimensional synthesis, the authors identify recurring design patterns, such as using RL to tune MPC parameters online or to replace the optimization layer entirely. Key challenges highlighted include computational burden, sample efficiency, robustness under model mismatch, and ensuring closed-loop stability guarantees. This work serves as a structured reference for researchers and engineers designing RL-MPC systems for linear control problems.
- Taxonomy covers 5 dimensions: RL roles, algorithms, MPC formulations, cost functions, and applications
- Identified challenges: computational burden, sample efficiency, robustness, and closed-loop guarantees
- RL commonly used for: online parameter tuning, policy optimization, and constraint handling in MPC
Why It Matters
This taxonomy provides a roadmap for engineers blending RL with MPC, cutting through fragmented literature.