Some concluding comments on Multi-Agent Systems

Chapter 06

Multi-Agent Systems (MASs) are still at a stage of development particularly in the domain of manufacturing systems. Continued advances are especially noticeable in the field of agent communication standards and improving agent development platforms. The most successful deployment of agent technology is in information industry. In this relatively new field (in comparison with manufacturing) difficulties accompanied with the introduction of new

technologies seem to be less prevalent. Firstly, because information systems represent a

homogenous environment (applications are entirely computational in nature like software agents are), and secondly, prohibitive traditional thinking and inertia are less manifested. To the contrary, however, the deployment of agent technology in traditional manufacturing systems has made much slower progress. Regardless of some early industrial implementations presented in research publications, in which agents have been moved out of the laboratory into the industrial workplace, it appears that a heterogeneous and more complex manufacturing environment, in which digital manipulations must be coupled with material transformation through elaborate control schemes, imposes many difficult problems for the deployment of agent technology. For example, most of the agent's scheduling architectures and protocols have only been developed to the theoretical stage. They are yet to be implemented and their viability and efficiency is yet to be tested. In addition, solutions to some of the difficulties that are fundamental to agent scheduling have not been found. For instance, although most researchers have put a lot of effort into achieving good communication among the agents, it is still a confusing and difficult problem. The first question to be answered is, "What information should each agent transfer in order to obtain high efficiency? Should we have a common information centre or should agents just provide information on their own status or should they make inquires about others in a point­ to-point fashion? Which is better, and in what situations? We may be able to find some answers in the general DAI (Distributed Artificial Intelligence) problems, but for manufacturing scheduling problems more specific answers are required. The second major question is how to prevent or solve the conflicts between the individual local solutions. As we learnt from studies and implementations of multi-agent scheduling systems, conflicts can be prevented by simply queuing up tasks (jobs) on critical resources and not redoing the overall scheduling at all. Routing flexibility has to be utilised as much as possible, and yet queuing cannot always be avoided. So far, in the cases where we do want to consider conflicts, there is no agreement on the proper way to do so. By "texture measures" [Sycara et al., 1 99 1 a] agents can foresee some of the possible conflicts, but again conflicts cannot be totally eliminated. "Negotiation" is, therefore, suggested by almost all authors. But these negotiations are far from mature in the sense that often they are just "doing negotiation for negotiation' s sake", that is, they only concentrate on solving conflicts, not on improving solutions.

In summary at the time when this project commenced, most of the above mentioned systems

were expensive. However the main reason why the development of the test bed application had not been conducted using the above tools was the lack of support in the case should something go wrong. In addition, it was considered that these tools were not yet sufficiently developed for practical application. Also many of the systems have proprietary features which do not make them suitable for universal application. On the other hand, it was expected that the Visual Basic

community of over 4 million users and Microsoft Corporation could provide prompt and

adequate support when needed. Also one of the project' s objectives was to use a general­

purpose programming language to avoid any proprietary issues that might be encountered during

the development.

Practical investigations as well as comparison with traditional techniques should be the focus of future work in this area. Only on the basis of such work it would be possible to find appropriate agent architectures and communication schemes for different real-world situations. It is for these reasons that the research in this project has concentrated on the development and implementation of an operating core of a heterarchical, Multi-Agent System, which could be used for comparing performance of such a multi-agent system with the other more conventional types of planning,

control, and scheduling systems described earlier in Chapter 3.

A review of manufacturing systems with particular problems in relation to manufacturing control systems indicated that modem small manufacturing organisations responding to rapid market changes and variable demand for their products posed particularly difficult scheduling problems. Many New Zealand manufacturing companies fall into this category making the solution of the scheduling and control systems particularly relevant.

In reviewing the control systems for manufacturing organisations such as those identified in this and the previous chapter, it appeared that heterarchical systems using agents should be able to provide a cost effective solution. At the same time it was clear that the newer technologies of the so called personal computers and networking had changed the cost effectiveness of this type of hardware and software, relative to the high cost of centralised computing systems, suggesting that distributed networks of the smaller computers could be utilised to provide the hardware and software basis for heterarchical, agent based distributed manufacturing control systems. From these deductions the research problem of interest for this study was developed to be as described in the following chapter.

Goran D. Colak _{Chapter 07}