Reinforcement Learning with Function Approximation for Manufacturing, Operations and Maintenance of Emerging Technologies

The optimization strategies of manufacturing, operations and maintenance are imperative for enhancing system quality, efficiency and reliability for many emerging technologies, e.g., lithium-ion batteries, high temperature superconductors, under stochastic operating environments and complicated manufacturing conditions. Many of these problems involve sequential decision-making in nature that can be modeled as Markov decision processes (MDPs). Recently, reinforcement learning (RL) has been implemented as an effective approach for solving MDPs in manufacturing problems. However, most RL algorithms applied to decision making are restricted to discrete state and action spaces. Continuous control variables are typically discretized to cope with the curse of dimensionality in traditional dynamic programming methods.

In this dissertation, we investigate the implementation of reinforcement learning throughout the product cycle (manufacturing, operations, and maintenance) of emerging technologies, where the problems are formulated into continuous state space, continuous action space MDPs by utilizing the function approximation. Different RL algorithms with function approximations are applied to solve these problems, while remaining computationally tractable. Computational studies of real data are conducted to demonstrate the advantages of the proposed RL frameworks in the manufacturing, operation, and maintenance problems.