Expert Systems in Decision Support

R.B.Jones, 82/11/1


3.1 The decline and fal1 of the mainframe

Three factors contribute to the projected decline in the profits which ICL can take from mainframw computers.

a) The market for medium to large computers is not expanding

The market for large computers is growing now much more slowly than that for minis and micros. This is partly due to rapidly improving technology which enables virtually all the epansion in processing power to be supplied with no increase in costs or revenues.

b) The proportion of the value of a large computer systems which resides in the central processor is declining.

Improvements in technology are more effective in reducing the cost of the processor than of the peripheral systems. Since ICL buys in mast of its peripherals this trend erodes the added value supplied by ICL and hence ICL profits.

c) The proportion of the value of a processor which is accounted for by the components is increasing.

As more and more components are packed onto each chip, the ratio of component costs to assembly and other costs is increaslng, Since ICL does not make components the value added by ICL is further eroded and profits fall.

3.2 The revival

Expert systems will I believe play a vital role among a number of factors which will during the next decade cause a reversal in the decline described above.

3.2.l The need for more processing power

Expert systems are open-ended mill consumers. This is because they typically involve searching through state spaces looking for (optimal) solutions to some problem. These state spaces would in the absence of good techniques for selecting avenues for exploration prove combinatorially intractable (i.e. take too long to search). Much of the research in Artificial Intelligence has been directed towards developing 'heuristics' which effectively prune the search space to manageable proportions. Nevertheless mill power remains an important factor in determining the Limitations of present day expert systems, and it is nost unlikely that in the forseeable future expert systems will fail to be thus limited.

The Japanese clearly believe that the expert systems of the future will be big consmers of mill. Their target for the power of the fifth generation computers ten years hence is 100 M Lips - 1 G Lips, one 'logical inference' takes between 100 and 1000 instructions on present architectures, so their target gives a machine which is as powerful (qua inference engine) as a 10 Gip Von Neumann machine. (taking the lower figure in both cases)

To some extent the mill requirements of expert systems can be expected to be distributed. Each individual user of expert decision support systems will expect a good proportion of his requirements to be supplied locally, which will itself provoke rapid increases in the power of personal workstations. This is only possible where the distributed system is supporting many users, each with relatively moderate problems. Big problems will be solved more economically on big processors, and large companies will find that such investments pay.

3.2.2 The diminishing importance of chip fabrication technology

For many years now the improvement in performance of processors has derived predominantly frcm improvaments in component technology. The significance of hardware architecture or of software has been relatively limited, in its impact on cost performance figures. The architecture of processors, mainfraine or micro, has not changed a great deal. This has enabled the design of microprocessors to be accomplished by quite small teams of engineers employed mainly by component manufacturers or companies with their own in house fabrication technology.

This trend has reduced the value added by manufacturers who are not vertically integrated, and has hence reduced their potential profits.

Though we can expect to see continuing rapid improvements in component technology, there are signs that this will be of dimishing significance by comparison with architectural problems and software problems. In addition, the monopoly which component manufacturers have had in the design of components shows signs of breaking up. ICL is already exploiting some of the first signs of this break-up in its use of Fujitsu gate arrays.

In the past custom VLSI chips have been designed in random logic by engineers with special knowledge of the technology. This enabled greater effective density to be achieived and hence resulted in more competitive products. The increasing complexity of VLSI circuits makes this sort of technique decreasingly viable, and progressively reduces the performance advantage obtained over more systematic computer aided design processes, as computer aided design becomes more extensively used for chip design, the relevance of specialist knowledge of the fabrication technology diminishes, and the design of the chip becomes largely a matter of logic rather than engineering. This opens the way for the 'silicon foundry' which fabricates silicon components to specifications supplied by the customer and processed by a design automation system. Such silicon foundries substantially reduce the disadvantage of not being vertically integrated. They enable companies such as ICL to have early access to technological advances, and they return to the computer manufacturer the design role.

This process is already well progressed. The Japanese fifth generation project thought entirely based on custom VLSI, will have no in house fabrication capability. All chips will be fabricated by subcontractors, who, provided they ,meet the quality standards, need not be Japanese. This means that ICL will have access to the same technology, and the best systen will be produced by he who best solves the architectural and software problem. As seen by the Japanese the problem is one of providing genuinely expert capabilities by the exploitation of vastly parallel architectures.

3.2.3 The increasing importance of architecture and software

The significance in the future of differences in hardware architecture (just how do we organise all this silicon?) and software engineering (knowlelge base organisation, inference techniques, heuristics, exploitation of hardware parallelism), will far outstrip the significance of the underlying fabrication technology. More of the factors determining the effective performance of processors will be within the domain of the computer designer rather than the component manufacturer. Automation in the design and production of components will enable much more use to be made of specialist hardware for different functions within the network. Hardware design will become less clearly distinct from software development and a greater flexibility about which functions are in silicon and which in software will result. All this is good news for systems designers, and increases the amount of added value contributed by the computer designer.

3.3 The strength of ICL

Some of the main areas of research in the Japanese fifth generation project are:

These are all areas in which ICL has quite well advanced products (DAP, CAFS and RDBMS, PERQ), and in which we can stay out front so long as we do not rest on our laurels. Substantial further development will be necessary to keep ahead in any of these areas if the Japanese come anywhere near acheiving their targets.

For example, DAP, though probably the most flexible parallel processing architecture commercially available, is nevertheless unlikely to be satisfactory for a parallel inferential capability. Although DAP is not contrained to parallelism in respect of floating point operations identically executed on large numbers of data items (as are most parallel processors), it is nevertheless constrained to doing pretty much the same thing on large amounts of data, be it floating point or not. This is a consequence of their being but the one program controlling simultaneously all the processors in the array. General problem, solving systems are unlikely to profit from this sort of architecture, requiring rather, an organisation in which each of the processors is more autonomous, allowing them to be allocated subproblems which are quite distinct. In such a context a heirarchic organisation of separately program controlled processors might be more appropriate than an array of commonly controlled processors. The Japanese appear intent upon stepping away from Von Neumann architectures altogether, in favour of 'data flow' architectures.

Similar advances will also be required in the other areas as well, it is unlikely to sufficient to produce cheaper functional replacements utilising more advanced component technology. It would be more appropriate to evaluate what more can be done with the new technology and revise the specifications upwards, this should however be done as part of a long term program which anticipates the rate of progress of the technology and aims at all times to 'intercept'. The importance of software developments in this process is likely to be considerable. The architectures of the future are likely to be unsuitable for human programming, and the ability to exoloit (and hence to sell) them will depend on having software expert enough to make use of them.

3.4 The Importance of Expert Systems

I submit therefore that 'expert system' are crucial to the marketing of large numbers of powerful processors in the following ways:

At the other end of the scale, increasing volume sales of less powerful systems will also depend upon the exploitation of expert systems techniques. This is because they will need to cope with people who simply don't understand computers, or people who's time is too valuable to be spent tailoring their problems to the computer system. They will want to explain their problems in the terms in which they understand them, and will require the responses to be immediately comprehensible.

up home © RBJ dated 1982/11/1 HTML 1996/6/6 edited 1997/4/18