Experiments with a distributed architecture for predictive scheduling and execution monitoring

Abstract

An important distinction is made in the planning domain between predictive planning problems and reactive execution problems [1]. The following paper discusses the distinction in the case of manufacturing scheduling. Given a set of jobs, an ordered set of production steps for each job, and a set of machines to perform these production steps, the predictive scheduling problems consists in determining an appropriate allocation of production steps to machines over time. Given partial or complete information about the current state of the manufacturing system, the reactive execution (or dispatching) problem consists in deciding “in real-time” which production steps to execute next on each machine. The paper presents an architecture allowing to associate a predictive scheduler and a reactive dispatcher. The predictive scheduler constructs a predictive schedule and updates it occasionally in response to arising problems and opportunities (in this sense, the predictive scheduler is a bit reactive!). The reactive dispatcher decides in real-time which operations to start on the shop floor given the actual course of events. It relies on the schedule to make its decisions, but overrides it in response to arising problems and opportunities. Our implementation allows the user to tune the behavior of both the predictive scheduler and the reactive dispatcher, and therefore to compare various couplings of predictive and reactive scheduling procedures. Experimental results provide evidence that independent settings of predictive scheduling and reactive execution algorithms can result in poor performance on the shop-floor.