Simulation is often used to evaluate supply chain or workshop management. This simulation task needs models, which are difficult to construct. The aim of this work is to reduce the complexity of a simulation model design. The proposed approach combines discrete and continuous approaches in order to construct speeder and simpler reduced model. The simulation model focuses on bottlenecks with a discrete approach according to the theory of constraints. The remaining of the workshop must be taken into account in order to describe how the bottlenecks are fed. It is modeled through a continuous approach thanks to a neural network. In particular, we use a multilayer perceptron. The structure of the network is determined by using a pruning procedure. For validation, this approach is applied to the modelisation of a sawmill workshop.
Simulation is used in many goals. One of them is to evaluate supply chain or workshop performance. There are three different ways of measuring this performance: analytical models (queuing theory, etc.), physical experimentation (lab platforms, industrial pilot implementation, etc.), and Monte Carlo methods (simulation or emulation) (Thierry et al., 2008). Analytical methods are generally impracticable because the mathematical model corresponding to a realistic case is often too complex to be solved, and physical experiments suffer from technical and cost-related limitations. Simulation is the better approach to model and analyze performance for large-scale cases. In the simulation model, the number of ‘objects’ of the model and the number of events can be very large. Consequently, the first problem could be the time needed to build the model and the simulation duration on a computer can be unacceptable for operational use. Thus, it is necessary to reduce the model size (Thierry et al., 2008).
On the one hand, constructing a simulation model is a complex task that can take modelers a lot of time. Effectively, simulation models of actual industrial cases are often very complex and the modelers encounter problems of scale (Page et al., 1999). Thus, numerous authors have expressed interest in using simplest (reduced/aggregated) models of simulation (Ward, 1989, Musselman, 1993, Pidd, 1996, Brooks and Tobias, 2000 and Chwif et al., 2006).
On the other hand, to establish and to initialize ‘predictive schedule’ or ‘reactive schedule’, the knowledge of the evolution of resources states (WIP (work in process) and queues) are needed. This knowledge can be obtained by using a simulation model. Reduced models can be very useful, because they are quickly parameterized and simulated.
A new approach for simulation model reduction has been presented here. This approach uses a neural network and, more particularly, a multilayer perceptron to model the functioning of a part of the process that is not constrained in capacity. This approach has been applied to the modelisation of a sawmill workshop. The results show that:
–
the two data sets present similar results,
–
the average of the error is small relative to the process time scale, and
–
the complete and reduced models gave similar results even if the log arrival rule is changed.
This means that it seems efficient to use a neural network to model a part of a process instead of constructing the complete model.
Assuming that the construction of a neural network is a quasi-automated task, in which the modeler only collects and selects the input data set. It is faster and easier to construct this kind of reduced model. This approach allows the modeler to focus on the management of bottlenecks.
Our intentions for future work are to investigate the structure determination of the neural network, particularly the choice of its inputs, and the validation of this approach on different applications, particularly on several external supply chains, such that at least one particular enterprise belongs to different supply chains.
