Proceedings of the 2013 Winter Simulation Conference R. Pasupathy, S.-H. Kim, A. Tolk, R. Hill, and M. E. Kuhl, eds INTELLIGENT DISPATCHING IN DYNAMIC STOCHASTIC JOB SHOPS Tao Zhang Oliver Rose Universität der Bundeswehr München Department of Computer Science D-85577 Neubiberg, GERMANY ABSTRACT Dispatching rules are common method to schedule jobs in practice. However, they consider only limited factors which influence the priority of jobs. This limited consideration narrows the rules’ scope of appli- cation. We develop a new hierarchical dispatching approach based on two types of factors: local factors and global factors, where each machine has its own dispatching rule setup. According to the global fac- tors, the dispatchers divide the state of the manufacturing system into several patterns, and parameterize a neural network for each pattern to map the relationships between the local factors and the priorities of jobs. When making decisions, the dispatchers determine which pattern the current state belongs to. Then the appropriate neural network computes priorities according to the jobs’ local factors. The job with the highest priority will be selected. Finally, the proposed approach is introduced on a manufacturing line and the performance is compared to classical dispatching rules. 1 INTRODUCTION In dynamic stochastic job shops, jobs are released and arrive at the shop over time. The release date, pro- cessing time of jobs and machine breakdowns are stochastic and not known in advance. It is difficult and sometimes impossible to compute optimal schedules. In this case, scheduling is typically carried out by means of dispatching decisions: once a machine becomes free, we decide what it should do next. Detailed dispatching decisions in a job shop are usually determined by dispatching rules. A dispatching rule can find a reasonably good solution in a relatively short time and is very simple to implement. In this paper, we assume that no machine is kept idle while a job is waiting for processing. At present, the study of dispatching rules focuses on two main fields: developing composite dispatch- ing rules and selecting rules dynamically. The first field aims to develop new dispatching rules which are applicable to more complex manufacturing lines. The rule never changes when being used. Holthaus and Rajendran (1997) develop five new dispatching rules of this type for scheduling a job shop. Some of these rules make use of the process time and work-content in the queue of the next operation of a job, by following a simple additive approach, in addition to the arrival time and dynamic slack of a job. The pro- posed rules are not only simple in structure, but also quite efficient in minimizing several measures of performance. Chen and Matis (2013) present a dispatching rule called the Weight Biased Modified RRrule that minimizes the mean tardiness of weighted jobs in an m-machine job shop. It is a significant extension of the RRrule in that it has linear complexity and considers weighted jobs. The shortcoming of these dispatching rules is that they are usually effective in some specific situations but not in others. Therefore, the second field focuses on finding better rules for a given situation. The rule changes over time according to the state of the manufacturing line. Scholz-Reiter et al. (2010) use Gaussian processes as a machine learning technique for the selection of dispatching rules in dynamic scenarios. Lian et al. (1998) propose a fuzzy inference-based selection of dispatching rules, adapting the scheduling decision to 2622 978-1-4799-2076-1/13/$31.00 ©2013 IEEE