Differentiable Projection Enforces Hard Constraints in Deep RL for Inventory Policies

A new arXiv preprint (2608.02343) introduces a method for learning feasible inventory policies in constrained sequential decision processes. The approach embeds a differentiable convex optimization module within a deep reinforcement learning (DRL) policy. A neural network proposes continuous action targets, which are then projected onto a relaxed feasible set via a quadratic program. A dual-informed integer mapping restores integrality while preserving feasibility. The policy is trained end-to-end using pathwise gradients from sampled trajectories in a differentiable simulator. This addresses limitations of existing DRL methods that either penalize constraints or rely on feasibility mechanisms that fail when constraints interact. The work targets operational problems with large, combinatorial action spaces and interdependent feasibility constraints, such as those modeled by mixed-integer linear programs (MILPs), which scale poorly in stochastic environments. The paper is announced as a new submission on arXiv.