Bridging Reinforcement Learning and Optimal Control via Feasible Action Mapping

What the math gives to ML

The paper offers a concrete safety mechanism: map an unconstrained RL action from a fixed abstract space into a state-dependent parameter of a feasible optimal-control problem before applying it. The transferable asset is not merely action clipping, but an optimization-based projection whose feasible set incorporates system dynamics, state/input constraints, and recursive feasibility. This can become a differentiable or partially differentiable safety layer between a neural policy and the environment, preserving a simple policy output space while the controller enforces constraints under the stated model and terminal-set assumptions. The key test is whether constraint violations remain zero in the certified regime and whether projection distance reveals the predicted feasibility boundary.