We study a production planning problem with linear, nonlinear, deterministic, and/or stochastic production relationships between the production plans and actual production quantities. We start by introducing a stochastic dynamic programming formulation of the problem with a Markovian assumption on the production relationships. Under specific conditions, we establish the convexity of the optimal cost-to-go function and closed forms of optimal policies. To solve the original problem in the general case, we propose a solution framework based on sequential policy optimization and deep reinforcement learning. We discuss the theoretical properties of the framework and evaluate its numerical performance with linear and nonlinear production relationships. In the linear case, our framework performs in line with the state-of-the-art optimization-based methods with improved computational efficiency. In the nonlinear case, our framework achieves an optimality gap near 10–20%. We also illustrate that the proposed methodology can also be extended to the problem of joint production planning and scheduling.
Liu et al. (Mon,) studied this question.