Frazier [89] investigated the effect of one-stage and  two- stage scheduling rules on different performance measures in a cellular manufacturing system. Fourteen scheduling rules and eight performance measures were used in the study. The simul- ation model was developed in SLAMII with FORTRAN sub- routines incorporated, and represented a  production  cell with six machines with separate queues for each part family. Two decision points were employed: the first one was switching between queues of part families or selecting the  next  part family queue, and the second one was selecting jobs in each part family queue. It was concluded that the MJ/SPT rule was ranked the best when all performance measures were con- sidered. The main shortcoming was that all performance meas- ures  were considered  to be  of  the  same importance.

Klein and Kim [90] demonstrated how a multi-attribute decision-making method could be used for dispatching  a vehicle. Using simulation models, the characteristics of decision rules for dispatching AGVs were shown. Single-attribute dis- patching rules consisted of LWT, STT/D, and MQS; and multi-attribute dispatching rules consisted of SAWM, YAGER, MMM, and MAWM methods. Performance measures con- sidered include average and maximum waiting time of a unit load in the output buffer, average  and  maximum queue length of the output buffer, job completion time, and total travel time of empty vehicles. However, no due-date-based performance measure was considered. The simulation model was tested for one, two, and three AGVs. Results showed that multi-attribute dispatching rules outperformed the single-attribute   ones.

Tung et al. [91] presented a hierarchical approach to schedul- ing FMSs with multiple performance objectives. The FMS consisted of a shop controller, a multiple-task sequencing con- troller with several AGVs, four CNC machines, one robot, an input buffer, an output buffer, and an inner buffer for storing

parts. The scheduling problem was solved at two  levels: the shop level, and the manufacturing system level. The shop level controller employed a combined priority index to rank shop production orders for meeting scheduling objectives. The FMS controller provided feedback to the shop controller with a set of suggested detailed schedules and projected order completion times. The shop controller then evaluated each candidate sched- ule using a multiple-objective function and selected the best schedule for execution. The proposed method was compared with SPT and EDD rules on four specific performance objec- tives, which were maximise profit, meet  due dates,  minimise the WIP inventory cost, and minimise finished-good inventory cost. Results indicated that the proposed scheduling method outperformed the two traditional scheduling rules. However, Tung  et al.  concluded  that  no  single  rule  was   universally the best.

4. Review of AI Scheduling  Approaches

4.1 Introduction

In a production system, the scheduling problem is to synchron- ise resources (connected by material transport system), and material flow, to produce a variety of parts in a certain period of time. Scheduling rules  are  used  to  select  the  next  part to be processed from a set of parts awaiting service. These rules can also be used to introduce workpieces into the system, to route parts in the system, and to assign parts to facilities such as workstations and AGVs. Some constraints also have to be considered  such as:

The schedule has to satisfy one or more  system  objectives, such as minimisation of mean flowtime or/and mean tardiness. Buffer  sizes  are limited.

Number  of transporters are limited,  etc.

Because of the complexity of the system, it is not very useful to find an optimal solution in an industrial  context  since changes often occur rapidly (for example, arrival of  new parts or modification of previous priority queue  size  of resources, and so on). Therefore, it is not desirable or economical when designing an optimal scheduler, but rather developing a flexible scheduling tool to assist the operator  to  monitor  the  system and make decisions. In fact, some operations can be replaced with an automatic scheduler  tool.  The  developed  tool  has to be easy to use, and react to changes in real-time. Consequently, it has to be expressed in terms of parameters that have to be chosen in accordance with the system objectives, which depend on the production situation. In the complex environment of an FMS, proper expertise and experience are needed for decision making. Artificial intelligence, together with simulation model- ling can help to imitate human expertise to schedule manufac- turing systems [92]. ElMaraghy and Ravi [26] reviewed some applications of knowledge-based simulation systems in the domain of FMSs, and also discussed their potential for the development of new, powerful and intelligent simulation environments for modelling and evaluating FMSs. Grabot and Geneste  [93]  stated  that  workshop  management  is  a    multi-