Relations between Model Predictive Control and Reinforcement Learning

The Evolution of Research in Microgrids Control

“…Comparison between MPC and RL: In comparison to optimal control methods, typically employing mathematical programming, like MPC, the biggest advantage of RL is its rapid online computation. RL is model-free, adaptive, with low online complexity, but contrary to MPC, its stability and feasibility are not guaranteed and its robustness is not backed by a solid theory, while it faces difficulties in handling constraints [112]. In general, an MPC agent is expected to outperform the corresponding RL implementation, if the identified model used is accurate.…”

Optimal Tracking Control for Uncertain Nonlinear Systems With Prescribed Performance via Critic-Only ADP

“…Violation of such constraints may cause that the parameter update process happens outside the feasible region, which makes no sense and leads to poor transient performance. However, the constraint handling abilities of ADP methods are still immature [33], and relevant studies are still very limited as mentioned in [13], [34]. Most of existing ADP controllers neither consider the estimation bounds of unknown parameters nor can meet the prescribed performance specifications (which can also be regarded as constraints) of system states.…”

Reinforcement Learning-Based Approximate Optimal Control for Attitude Reorientation Under State Constraints

“…It can iteratively approximate the optimal control policy and avoid analytically solving the intractable HJB equation. However, the state constraint handling ability of RL-based control is still immature [23]. In this paper, a constrained RL-based control method is proposed, and a critic-only neural network (NN) structure is designed to approximate the optimal cost function and control policy.…”