Search papers, labs, and topics across Lattice.
This paper introduces a novel approach that integrates a differentiable convex optimization module within a deep reinforcement learning (DRL) policy to effectively manage hard constraints in combinatorial action spaces. By allowing a neural network to propose continuous action targets that are then projected onto a feasible set, the method maintains feasibility while achieving high scalability in stochastic environments. The results demonstrate that this approach not only achieves an optimality gap below 1% in small instances but also outperforms existing policies by significant margins in larger, more complex production-inventory planning scenarios.
Achieving less than 1% optimality gap while outperforming state-of-the-art policies by up to 9.75% reveals a breakthrough in managing hard constraints within DRL frameworks.
Many operational problems are constrained sequential decision processes with large, combinatorial action spaces and interdependent feasibility constraints. Mixed-integer linear programs (MILPs) handle such constraints flexibly but scale poorly in stochastic environments. Deep reinforcement learning (DRL) promises scalable decision rules, but existing methods either penalize constraints rather than enforce them, or rely on feasibility mechanisms that break down once constraints interact. We bridge this gap by embedding a differentiable convex optimization module inside the policy: a neural network proposes continuous action targets, a quadratic program projects them onto the relaxed feasible set, and a dual-informed integer mapping restores integrality while preserving feasibility. Given a differentiable simulator, the policy trains end to end from sampled trajectories using pathwise gradients, while handling hard constraints with similar flexibility to MILPs. We show that our feasibility enforcement has bounded error relative to an exact integer projection and ensures the entire feasible action space is reachable. We apply the method to multi-echelon production-inventory planning under shared resource and material constraints. Our policy attains an average optimality gap below 1% on small instances. It further outperforms state-of-the-art echelon base-stock policies by up to 9.75% and a rolling-horizon multi-stage stochastic program by at least 7.7% in larger networks. On an industry-scale case study from ASML, it reduces average cost by up to 3.22% relative to the best-known benchmark policy. The savings are largest where planning is hardest: in tightly capacitated systems with high demand variability. More broadly, our work shows that DRL can deliver economically significant savings in sequential decision problems with interdependent hard constraints, which are widespread in practice.
