In the past decades, many optimization methods have been devised and applied to job shop scheduling problem (JSSP) to find the optimal solution. Many methods assumed that the scheduling results were applied to static environments, but the whole environments in the real world are always dynamic. Moreover, many unexpected events such as machine breakdowns and material problems may be present to adversely affect the initial job scheduling. This work views JSSP as a sequential decision making problem and proposes to use deep reinforcement learning to cope with this problem. The combination of deep learning and reinforcement learning avoids handcraft features as used in traditional reinforcement learning, and it is expected that the combination will make the whole learning phase more efficient. Our proposed model comprises actor network and critic network, both including convolution layers and fully connected layer. Actor network agent learns how to behave in different situations, while critic network helps agent evaluate the value of statement then return to actor network. This work proposes a parallel training method, combining asynchronous update as well as deep deterministic policy gradient (DDPG), to train the model. The whole network is trained with parallel training on a multi-agent environment and different simple dispatching rules are considered as actions. We evaluate our proposed model on more than ten instances that are present in a famous benchmark problem library – OR library. The evaluation results indicate that our method is comparative in static JSSP benchmark problems, and achieves a good balance between makespan and execution time in dynamic environments. Scheduling score of our method is 91.12% in static JSSP benchmark problems, and 80.78% in dynamic environments.

The dynamic job-shop scheduling problem (DJSSP) is an advanced form of the classical job-shop scheduling problem (JSSP), incorporating dynamic events that make it even more challenging. This article proposes a novel approach involving deep reinforcement learning and graph neural networks to solve this optimization problem. To effectively model DJSSP, we use a disjunctive graph, designing specific node features that reflect the unique characteristics of JSSP with machine breakdowns and stochastic job arrivals. Our proposed method can dynamically adapt to the occurrence of disruptions, ensuring that it accurately reflects the current state of the environment. Furthermore, we use the attention mechanism to prioritize crucial nodes while discarding irrelevant ones. This study proposes a model that applies graph attention networks to learn node embeddings, serving as input for the actor–critic model. The proximal policy optimization is then utilized to train the actor–critic model, which assists the model in learning the scheduling of job operations for machines. We conducted extensive experiments in static and public environments. Experimental results indicate that our method is superior to current state-of-the-art methods.

Parallel machine scheduling (PMS) is a common setting in many manufacturing facilities, in which each job is allowed to be processed on one of the machines of the same type. It involves scheduling $n$ jobs on $m$ machines to minimize certain objective functions. For preemptive scheduling, most problems are not only NP-hard but also difficult in practice. Moreover, many unexpected events, such as machine failure and requirement change, are inevitable in the practical production process, meaning that rescheduling is required for static scheduling methods. Deep reinforcement learning (DRL), which combines deep learning and reinforcement learning, has achieved promising results in several domains and has shown the potential to solve large Markov decision process (MDP) optimization tasks. Moreover, PMS problems can be formulated as an MDP problem, inspiring us to devise a DRL method to deal with PMS problems in a dynamic environment. We develop a novel DRL-based PMS method, called DPMS, in which the developed model considers the characteristics of PMS to design states and the reward. The actions involve dispatching rules, so DPMS can be considered a meta-dispatching-rule system that can efficiently select a sequence of dispatching rules based on the current environment or unexpected events. The experimental results demonstrate that DPMS can yield promising results in a dynamic environment by learning from the interactions between the agent and the environment. Furthermore, we conduct extensive experiments to analyze DPMS in the context of developing a DRL to deal with dynamic PMS problems.

Predicting the wafer material removal rate (MRR) is an important step in semiconductor manufacturing for total quality control. This work proposes a deep learning model called a fusion network to predict the MRR, in which we consider separating features into shallow and deep features and use the characteristics of deep learning to perform a fusion of these two kinds of features. In the proposed model, the deep features go through a sequence of nonlinear transformations and the goal is to learn the complex interactions among the features to obtain the deep feature embeddings. Additionally, the proposed method is flexible and can incorporate domain knowledge into the model by encoding the knowledge as shallow features. Once the learning of deep features is completed, the proposed model uses the shallow features and the learned deep feature embeddings to obtain new features for the subsequent layers. This work performs experiments on a dataset from the 2016 Prognostics and Health Management Data Challenge. The experimental results show that the proposed model outperforms the competition winner and three ensemble learning methods. The proposed method is a single model, whereas the comparison methods are ensemble models. Besides the experimental results, we conduct extensive experiments to analyze the proposed method.