The development of techniques to select a control policy during proactive on-line planning and control

[摘要] ENGLISH ABSTRACT: The worldwide trend for systems is to become more complex. This leads to the need for newways to control these complex systems. A relatively new approach for controlling systems, calledon-line planning and control, poses many potential benefits to a variety of end-users, especially inthe manufacturing environment. Davis [3J developed a framework for on-line planning andcontrol that is currently incomplete. This project aims to fill one of the gaps in the framework byautomating one of the functions, eliminating the need for a human observer. This function, thereal-time compromise analysis function, does the comparison of the statistical performanceestimates to select a control policy for implementation in the system being controlled (the realworldsystem) at the current moment in time.In this project, two techniques were developed to automate the function. The first technique isbased on a common technique for statistically comparing two systems, the paired-t confidenceinterval technique. The paired-t confidence interval technique is used to compare the controlpolicies by building confidence intervals of the expected differences for the respectiveperformance criteria and testing the hypothesis that the statistical performance estimates of theone control policy are better than those of the other control policy. The results of thesecomparisons are then consolidated into a compromise function that is used to determine thecontrol policy to be implemented currently in the real-world system.The second developed technique is derived, but differs greatly, from Davis's [3J dominanceprobability density function approach, and it includes principles of the paired-t confidenceinterval technique. It compares the control policies by determining the probability (confidencelevel) with which one can assume that the performance criterion of the one control policy willprovide a performance value that is better than the other's and vie-ursa. These confidence levelsare then aggregated into a single compromise function that is used to determine the controlpolicy to be implemented currently in the real-world system.After the techniques were developed, it was not possible to determine their efficiencymathematically, because their statistical base is suspect. The techniques needed to beimplemented before they could be evaluated and it was decided to develop an emulator of theon-line planning and control process in accordance with the framework given by Davis [3J toimplement them. This Emulator is in essence a Visual Basic program that uses Arena models.However, this Emulator needed certain deviations from the framework to make it possible.Firstly, while the systems that will be controlled with the on-line planning and control processwill be complex systems, the system controlled in the Emulator is only a straightforwardMlM/l/FIFO/OO / 00 system. This allowed for the conditions that have not been addressedsufficiently, e.g. the initialising of the system models, to be bypassed. Secondly, the Emulatordoes not include all parts of the framework, and parts for which the technology does notcurrently exist have been excluded. Thirdly, the real-world system is replaced with a model,because a real-world system was not available for the study. Finally, concurrent operations areactually done sequentially, but in a way that makes it seem that they were done concurrently, asnot to influence the results.This Emulator was used to analyse both techniques for two different traffic intensities. The firstpart of the analysis consisted of an off-line non-terminating analysis of the individual controlpolicies of the system. This was used as a base line against which the on-line planning and controlprocess of the Emulator was evaluated.The findings of the evaluations were that, at the traffic intensities evaluated, the techniquesprovided results that were very similar to the results of the best individual control. From theseresults, it was speculated that at different traffic intensities, different control policies would bebetter than the techniques themselves, while the techniques will only give slightly worse results.In addition, because the on-line planning and control process attempts to respond to changingconditions, it can be assumed that the techniques will excel in those conditions where the inputdistribution is changing continuously. It is also speculated that the techniques may beadvantageous in cases where it is not possible to determine beforehand which of the individualcontrol policies to use because it is impossible to predict the input distribution that will occur. Itis expected that the techniques will give good (but unfortunately, not necessarily the best) resultsfor any input distribution, while an individual control policy that may give the best results for oneinput distribution, may prove disastrous for another input distribution.Three important conclusions can be made from the project. Firstly, it is possible to automate thereal-time compromise analysis function. Secondly, an emulator can be developed to evaluate thetechniques for the real-time compromise analysis. The greatest advantage of this Emulator is thatit can run significantly faster than real-time, enabling the generation of enough data to make thesignificant statistical comparisons needed to evaluate the techniques. The final conclusion is thatwhile initial evaluations are inconclusive, it can be shown that the techniques warrant furtherstudy.Three important recommendations cart be made from the project. Firstly, the techniques need tobe studied further, because they cannot be claimed to be perfect, or that they are the onlypossible techniques that will work. In fact, they are merely techniques that may work and othertechniques may still prove to be better. Secondly, because it would be foolhardy to assume thatthe Emulator is complete, the Emulator needs to be improved with the most critical need todevelop the Emulator in a programming language and simulation package that allows concurrentoperations and effortless initialisation. This will enable the Emulator to be much faster and a lotmore flexible. The final recommendation is that the techniques need to be evaluated with otherparameters in other increasingly complex systems, culminating in the evaluation of the on-lineplanning and control process with the techniques included in a real-world flexible manufacturingsystem. Only then can there be decided conclusively on whether the techniques are efficient ornot.It is hoped that this project will form a valuable building block that will facilitate making on-lineplanning and control a viable alternative to controlling complex systems, enabling them torespond better to changing conditions that are currently becoming the norm.

[发布机构] Stellenbosch University