Stochastic Optimization and Deep Reinforcement Learning for Real-Time Decision-Making in High-Variability Manufacturing Systems focus on uncertainty- demand changes, machine breakdown

Abstract

Modern manufacturing systems operate under unprecedented levels of uncertainty, including stochastic demand fluctuations, unpredictable machine breakdowns, and supply chain disruptions. Traditional deterministic optimization and rule-based control methods often fail to adapt in real time, leading to suboptimal throughput, high energy costs, and schedule infeasibility. This paper proposes a hybrid framework integrating Stochastic Optimization (SO) for scenario-based pre-positioning and Deep Reinforcement Learning (DRL) for real-time policy execution. We model a high-variability job shop as a Markov Decision Process (MDP) with hidden and stochastic transitions. A DRL agent, specifically a Proximal Policy Optimization (PPO) algorithm augmented with a risk-sensitive objective, learns adaptive scheduling and maintenance-triggering policies. Comparative simulations against dispatching rules (SPT, CR) and model predictive control (MPC) demonstrate that the proposed framework reduces average job tardiness by 34%, improves machine utilization by 18% under breakdown conditions, and maintains robust performance under demand volatility up to 40% coefficient of variation. The paper concludes with a discussion on real-time deployment architectures and future integration of digital twins.