The application of inductive learning in simulation of queuing systems.

摘要

Expert systems may be used for enhancing the applicability of computer simulation as a decision support tool. This could assist in the process of simulation model modification and refinement which would achieve a set performance goals for the system under study. Machine learning can be applied as an effective method for extracting heuristic rules to support the knowledge base of such an expert system.; In this research, the effectiveness of inductive learning, more specifically the family of ID3 algorithms, was studied to obtain the required rules. This tool was tailored for output analysis of a simulated queuing system. Several examples of the system, referred to as instance-sets, are required for learning. Each instance is represented by its class and by several control-features of the system.; This research initially focused on control-feature engineering in queuing system simulation. Later, the focus was on artificial instance generation for the required instance-set. It is shown that the generation of the required instance-set is a complex search problem. An automatic instance generation procedure is proposed which assists in generating suitable instance-sets in the absence of realistic instances. The proposed procedure is a combination of the three search methods of grid base, forward search, and backward search.; Experiments were initially performed on those examples of a M/M/1 queuing system for which enough information to validate the experiment progress and final results was available. Later, experiments with a two-serial-server queuing system are used to generalize the findings to some extent.; This report shows that in general the induction tree algorithm generated prediction-rules that were effective in the output analysis of the queuing systems under study. Also, the designed control-features for the queuing system indicated no preference over the common control-features. The proposed automatic instance generation procedure has been suitable for this research, and it could be updated to match similar situations. The results further suggest that inductive learning can be considered as a suitable alternative to knowledge acquisition for expert systems that are used in the analysis of simulation output.