Integrate across syndicates with respect to:-
£100K, (SERC/DoI), October 1982 - April 1983
Prior to Main IKBS SPP and ALVEY IKBS project:
include:-
This talk summarises the proposals put forward to the Alvey Committee by Charles Read of the Committee and myself, based on wide information gathering from the academic and industrial community.
Essentially, our proposals were for a consolidated, larger SPP, with more participation by industry and more reference to other areas of IT. We thought of the SPP activity as feeding into and drawing from wider IKBS and IT communities. The detailed Working Paper we prepared could thus be regarded as an amplification of the SERC's SPP document.
We characterised IKBS as limited intelligence systems. They are exemplified by systems for giving initial instruction in programming, or for summarising news stories. IKBS are intermediate between the unintelligent systems of the present and the fully intelligent systems of the remote future. The former are systems for executing a single task with a complete specification, the latter systems for executing several tasks with incomplete specifications. An IKBS is a system for a single task, but one with an incomplete specification. Further, the complexity of the task implies working with large and/or incomplete, uncertain, rapidly-changing knowledge stores; with tentative and/or heterogeneous inference procedures for applying knowledge in response to varied and/or unreliable inputs in changing environments; and with multiple and/or flexible control mechanisms.
This characterisation of IKBS was adopted as useful in emphasising the application of intelligence as inferential problem solving relying on non-trivial knowledge, i.e. as defining an IKBS as a system using inference to apply knowledge to perform a task. The characterisation was also seen as allowing wide scope for different types of system, and for a range of systems from the humble ones the community is attempting to build now to the more ambitious ones it would like to build in the future. (We treated expert systems as a subset of IKBS characterised by specialised knowledge and informed users.)
Our description of the structure of an IKBS was seen primarily as a means for identifying what is available for building IKBS, what is needed, and what can be hoped for, and hence as providing a basis for organising a programme of R&D in IKBS. We adopted the idea of levels of knowledge representation and problem solving put forward in the SPP document, but expanded to four levels. These are the technology or formal processing level e.g. of a reduction machine; the capability or basic processing level, covering both internal operations e.g. of planning and external operations e.g. of object sensing; the task or generic processing level e.g. of scientific diagnosis; and the application or domain processing level e.g. of circuit layout. These levels are all involved in a single system, as in e.g. a Lisp technology realisation of a natural language capability for a database questioning task for a high energy physics application. The problem of building IKBS is that of identifying what is needed at each level and of integrating across levels. This can be thought of as defining and relating processes and their corresponding processors, possibly, as in a robot, of very different kinds, i.e. both general and special, high level and low level, internal and external. Building an IKBS is also subject to the constraints of the system's social context e.g. user convenience, its engineering context e.g. real time limits, and its computational context e.g. machine reliability.
To provide a starting point for the proposed programme, we looked at what is available at the individual system levels and at what is known about system building as a whole. We found some concrete tools at the technology level e.g. Prolog; at the capability level e.g. semantic nets, stereopsis algorithms; at the task level e.g. pupil modelling; and at the application level e.g. mass spectra characterisation. We also noted the concrete experience in system building represented by e.g. Prospector, R1, Intellect. But it is evident that there are many serious problems and hence needs for research at the levels e.g. at the technology level into robot languages, at the capability level into rule induction and robust parsing, at the task level into teaching strategies, and at the application level into metal welding (i.e. into the characterisation of this as something to be intelligently automated). There are equally many problems and hence needs for research in overall system design e.g. into combining formal mathematics with informal object notions in image interpretation. These needs are both for the development of concepts and for the provision of utilities. They are very great, but we concluded that some effective leverage can be obtained from what is already known.
We thought it important, in looking at the requirements and possibilities of IKBS, not to be too narrow, but to consider the wider relations between IKBS and other areas of IT. We noted mutual relations of supply and demand, both direct and indirect, with IKBS actually or potentially supplying intelligence and demanding (good) unintelligence. There are direct relations between IKBS and the man-machine interface area, with IKBS supplying intelligent interfaces and demanding other interface technology; between IKBS and software systems, with IKBS supplying intelligent system components and demanding other components; and between IKBS and hardware/VLSI, with IKBS supplying intelligent operations and demanding other operations. There are also indirect relations, especially with software engineering, through the supply of IKBS design tools and the demand for regular design tools, but also elsewhere e.g. through the supply of IKBS for VLSI design. Some relations were especially considered for Alvey purposes, but were regarded as selected illustrations of the general forcing function of IKBS work for computing as a whole, i.e. of IKBS promoting and wanting better computing system tools.
Putting our characterisation of IKBS together with the community's mixed ignorance and knowledge of IKBS, in the context of the importance and utility of IKBS in IT, led us to propose a programme for IKBS like the SPP, with three major elements, namely research programmes, demonstrator projects, and education and training. These components reflect the needs of the area for long-term research on system properties at level sand overall, for shorter-term development of state of the art systems, and for a substantial increase in manpower to carry out the programmes and projects. But the programme is on a larger scale than the SERC SPP, with the major addition of purely industrial research programmes as well as academic and joint programmes. The addition is a response to the importance and variety of industrial interests in IKBS relating both to industry's internal efficiency and to its market products, and so to its concerns with IKBS as system researchers, producers, and consumers. We noted that both the IT industry and IT-using industry can be any of these, and that both large and small companies can be in vol ved. We regarded industrial participation in the programme as especially important in emphasising the ultimate market uses of IKBS, which are particularly critical for successful demonstrator projects.
We thought it important, further, to make the point that there are two special challenges for the proposed programme, beyond those normal for any long-term programme of research and advanced development. These are the manpower requirements for the U.K., and the intrinsic difficulty of the field. We therefore stressed the need to build up the community, and to support some really basic research. We further stressed the need, even with a successful programme, to be realistic about what can be achieved. It is essential not to oversell IKBS. However our belief is that systems worthy of the name IKBS can be built in the U.K., starting now.
To give substance to our proposals, we considered what we should aim for, in the sense that we can expect to attain it, especially in the first five years of the programme where it is possible to be fairly definite.
We claimed that in 2-3 years we can expect a better understanding of capabilities e.g. frame application, speech recognition; of tasks e.g. database querying; and of applications e.g. equipment checking. In 5 years we can expect further understanding of capabilities e.g. text topic identification; of tasks e.g. program drafting; and of applications e.g. process control. We can also expect substantial progress in understanding IKBS as wholes, and of how to integrate the modules of a system. We can expect significant advances in relevant technologies both conventional and non-conventional, especially in logic programming and the use of parallel processors. And we can expect some provision of utilities e.g. parsers, vision processors, perhaps in some cases in 2-3 years. In 10 years we can expect further progress to have been achieved, but detailed examples cannot be given now.
In 2-3 years we claimed we can expect a range of modest expert systems, along with some rather more ambitious IKBS e.g.
In 5 years time we can expect some more ambitious systems e.g.
In 10 years we can expect that even the most difficult demonstrators could have been implemented, if only in a modest form, and in general that progress in research should have resulted in IKBS demonstrators significantly more powerful than anything available today. Some of these would be achieved by incremental improvement e.g. citizen's helper with voice, others by radical breakthroughs, as would be required eg for a robot with full 3D perception.