This document represents a hurried compilation of the syndicate discussions. The major part of our Syndicate meeting was occupied by 14 individual presentations. A number of interests were covered such as: databases, logic, fuzzy sets, VISI design, software design and engineering applications. A lot of interest in the use of PROLOG was expressed although people were mindful of its shortcomings. Many presentations referred to teams like knowledge or knowledge representation but no precise definitions were offered although many viewpoints were evident.
Central to the discussions was an attempt to arrive at a consensus, or more precisely, a shared intuition about knowledge representation. The following points emerged:
The discussion led us to identify in the first instance four general areas for attention in research programmes. These are:
We felt that it was too soon to establish firm overall objectives and timescales in this vital area of research but it was important to acquire the know-how as soon as possible. Some of the questions that the research proposals should attempt to answer on the subject of representation are:
Everyone present was interested in these topics and indeed some work was already under way at several centres.
Requirements from other areas: Knowledge representation relates to every other area within the IKBS initiative discussed at this meeting. It has particular relevance to the software tools and software methodology within the total Computer Science discipline.
This is by no means a complete list of interests shown at the meeting. Other research interests were expressed but not formulated.
List of necessary support facility: There was much concern about the hardware and software environment. No detailed discussion took place. It was left to the members attending the infrastructure group to address this issue. It was felt that the demand for facilities should be based not simply on present technology (VAX, PERQ, LISP machine) but what future technology can supply (large and cheep memories etc).
Demonstration Projects: There was strong agreement that demonstrator projects should be designed to be attractive to industry. For this to be so, they must identify a clearly felt industrial problem area.
Following projects were suggested:
One demonstrator project on which interested parties have put in a considerable thought was proposed. A possible title for this is:
Knowledge based System for Management Decision Support for CAD/CAM tasks.
The proposal originates from the following parties:
Following other parties are also keen to participate in this project:
Another potential demonstrator project currently under way is:
Software Engineering Workstation Environment,
on which considerable planning has been carried out by:
It was evident from the discussion that industrial representatives in the syndicate were keen to participate in the design and implementation of demonstrator project.
Rapporteur: E H Mamdani
Chairman: K Baker
We regard inference as one of the core subjects of basic artificial intelligence (other core subjects include knowledge representation and planning). These subjects tend to be characterised by being extremely difficult. Although the past few years have seen increasing utilisation of existing techniques (for example: the increasing use of Prolog; the exploitation of new technology; the realisation of the concept of computation as controlled deduction etc) there has been very little vital research into inference - no major conceptual advances have been recognised.
Although inference techniques needed for IKBS in practice are currently fairly shallow, and are likely to remain so in the short/medium term (up to five years), in the longer term new techniques will be needed which will depend on basic research being carried out now. Indeed, one possible outcome of this research should be a fuller understanding of which parts of intelligent systems may properly be called inference. Lack of definition in this respect made our task of selecting possible project areas somewhat difficult.
The following list of possible projects is not intended to be exhaustive; projects arising from individual proposals should not be excluded. We feel that the SPP as a whole will need very careful management; the expansion of research in this area will yield a requirement for additional assessment resources. We should ensure that the review procedures will be a s careful as the normal SERC peer review system.
In addition, some of the following projects have names of interested parties associated with them. Again, these lists of names are not intended to be exhaustive; they are examples of people or groups who are interested in the appropriate areas and are known to members of the syndicate.
These projects can be applied to pattern recognition.
Some of the preceding projects can be construed as or can yield demonstrator projects.
For example:
Inference is closely related to all areas of IKBS. It is particularly close to knowledge representation. Some specific relations are indicated under the individual project areas; we emphasize again that inference is essential to any IKBS.
We anticipate sufficient resources for the academic projects coming from the SERC common base.
Henry Thompson - Rapporteur
A crucial commitment of any programme in IT must be to making computational systems accessible to and useable by a wide range of users. In both the short and long term, natural language has a key role to play in bringing this about.
As per the remit of the syndicate, there follows hereafter an outline of major long-term research areas, which we feel are of major importance, together with associated tools and facilities, and a number of near-term development projects which might be undertaken to demonstrate the state of the art in various areas.
In considering short-term development projects and long-term research goals alike, it is helpful to hold one distinction in mind, between knowledge-rich and knowledge-free language processing. That is world knowledge, needless to say. Knowledge-free systems operate on a linguistic basis alone, and are thus free to address problems which are unrestricted as to content and domain. Machine Translation is the paradigm task here, along with document retrieval and, from another perspective, text to speech systems such as those in reading machines for the blind. We identify two research areas here:
The first of these is directed at examining large quantities of natural English text, to identify end quantify those superficial relationships which exist therein. These relationships can then help in a number of ways in large scale natura1 language processing.
The second is concerned with carrying such text into a sort of normal form, which can serve as the basis for translation, summarisation, search, etc.
Knowledge-rich processing, on the other hand, depends crucially on domain specific, non-linguistic knowledge. As such it is closely tied to concerns in Knowledge Representation and Inferencing, and indeed the attempt to distinguish between these 3 labels in this area is often counter-productive. We identify 3 research areas here:
The first of these is concerned with the theoretical basis needed to make the necessary move from sentences to dialogues and texts as our domain of concern. Most communication seems directed at the co-ordinated influencing of what others believe and desire, and of what they intend doing about relevant aspect of those beliefs and desires.
We need to develop models of beliefs, desires, and plans which are rich enough to encompass the subtlety and power of natural language use. The other two research areas are obviously closely tied to this one. Language analysers to date have been fragile in large part because they lacked such a rich contextual structure. Analysis systems incorporating a broad motivational model should be able to cope with a much wider range of input ungrammatical, fragmentary, imperfect in a number of ways. Likewise for the complex, knowledge intensive systems of the future to be able to explain their own operation to their users, as they must be able to do, once again a rich model of both user's and machine's beliefs and goals is required.
And finally an area of research that resists, or rather spans, the knowledge-rich/knowledge-free dichotomy:
From the details of signal processing and acoustic phonetics, through the higher level issues of system organisation, this is vast and complex task. The state of the art in connected word recognisers represents something of a plateau, progress from which will be difficult, but there is clearly much that can be done.
In closing the section on research areas, it should be emphasised that the knowledge-rich/knowledge-free distinction is an overly facile one, used here as an aid to organisation. Those terms name only the extremes of a continuum and most issues of concern find some manifestation at almost any point along it.
The two major theoretical ties should be clear from the above - a pervasive inter-dependence with linguistics, and, as one gets closer to the knowledge-rich end, with psycho-linguistics, knowledge representation, and inferencing. Indeed in the limit the question is the same: How do we get the world into the machine in a useable way and then use it?
These fall into 2 broad categories - specialised tools from within the natural language area, and general computational support, both hardware and software. In the first category a number of well understood sub-systems at the capability level need to be enhanced, packaged, polished and made available to the community:
All of these are resources needed both by the research projects above and in the development projects given below.
As for computational facilities, our needs are similar to others. and the IKBS common base is a good starting point. But NL processing is particularly demanding of both address space and cycles, and 3 VAXes would hardly satisfy our needs, much less leave anything for anyone else. We need a network of VAXes and suitable smaller machines with a good range of languages (LISP, POP, PROLOG, BCPL/C/PASCAL), a good range of support software (editors, debuggers, text formatters) and a commitment to the provision of adequate support staff on a local and/or regional level.
We have identified 4 projects which might serve to demonstrate where we are and can easily get to in the short term.
There is much in common between these three, both in terms of common facilities employed, in starting out relatively knowledge-free with room to grow in the knowledge-rich direction, and in all being focused on language itself both as medium and as object.
We also feel the time is right to mount a significant demonstration in the area of highest visibility of NL:
We confess to a certain scepticism as to the appropriateness of natural language to this task, especially on the input side, but feel the time is right for a high-quality experiment to address the issue. We envision a 3-stage project - the first stage using existing technology on large machines; and 2 further stages in parallel, one moving towards small size and low cost, and the other towards speech input/output, both using the machinery referred to above under tools.
Help! There is a dramatic shortfall in the necessary skilled manpower. We need research studentships in substantial numbers, targeted to particular projects and/or members of staff, with special emphasis on encouraging recruitment from relevant non-computational fields, including a 4th year of support for such students. Consideration should be given to establishing a P-G course specifically for training computational linguists. How we are to simultaneously teach the necessary people and do the necessary research, to say nothing of guiding demonstration projects, is more than we can say.(There is also a clear communication problem with industry and efforts are needed to overcome this.)
In closing, we wish to emphasise that although everyone's image of the millennial IT system crucially involves natural language, we cannot now produce such systems, nor can we even see how to get there from here. There is a need for research on a broad front, with a long term horizon, to lay the foundations on which such systems may hope to be built.
(J P Frisby - Rapporteur; A Blake; J E W Mayhew)
The Syndicate was of small size and unrepresentative of the UK vision community. If a planning exercise for vision research is to be done seriously, SERC will have to launch a much more thorough and lengthy review. Also, the Syndicate supports the continued use of the peer review system for any detailed research programmes that emerge.
The past few years have seen the development of good theories of various low level vision tasks, such as stereo, motion parallax, line funding. The time is now ripe to exploit these theories in algorithms capable of being run in (near) real-time, perhaps using parallel or other special architectures.
Considerable current interest is being shown in the exploration of algorithms suitable for implementation on parallel architecture (eg for 2D image analysis). The goal here is to characterise those algorithms capable of exploiting the potential benefits of parallel processors. Obviously, this area might overlap with Programme A above. although the initia1 starting points of A and B are different. A-type work starts from a theory and asks: how can it be implemented to work quickly, perhaps on a parallel machine? B-type work starts by defining algorithms which run efficiently on paral1el array machines, and investigates their scope.
Vision presents an excellent domain for fundamental research on intelligence. It presents a range of highly complex information processing problems, such as the analysis of complex 3D scenes, which can now be studied from the valuable vantage point of solid achievements in low-level vision. The IKBS SPP ought to find room to support such basic research, even though it is high risk and unlikely to present practical benefits in the short-term.
The 2.5 D Sketch is a term coined by Marr for a viewer-centred representation of the visible surfaces of a 3D scene. It is a dynamic memory buffer integrating the information delivered by multiple lower-level modules, such as stereo, shape-from-shading, etc. Its computation is a highly non-trivial problem, presenting fundamental questions about the details of a suitable representation can access it. In a sense, the 2.5 D sketch is the culmination of all that could be termed low-level visual competence. It is the obvious database for the computation of 3D object representations, and hence provides the natural, albeit large, next step forward.
There appears to be a need for basic research on the learned, as opposed to programmed, acquisition of visual capability, ie research into visual systems capable of learning as they accomplish a task. Current examples are inductive pattern classification systems of the kind built by Michie's group in Edinburgh, and the large parallel memory networks being studied by Anderson, Alexander, Hinton in the United States.
The report of the Syndicate on Object Manipulation specifies many needs from other areas which would obviously benefit vision research. In addition, we would particularly draw attention to parallel hardware developments.
The machine of choice at present for a vision group is a medium-sized VAX running UNIX. If proposed developments to the PERQ are realised (eg cost reduction, I/O improvements, UNIX, etc) then it may serve as a good vision workstation.
It is essential to have fast image acquisition and display facilities. Adequate technical support staff are crucially important.
Vision work in the UK is fragmented into small groups working in isolation, each lacking the critical mass characteristic of their USA counterparts. It is doubtful that this situation can be fundamentally changed (eg by the setting up of a large institute), in which case it is essential that adequate travel and conference support should be made available to help integrate research efforts.
Funding for attendance at major international conferences is absolutely essential. Failure to attend such meetings can slow progress enormously and worse, lead to work being undertaken that has already been completed elsewhere.
SERC should also consider setting up a working party to support the integration of the work of different groups.
The Syndicate was unable to arrive at a well-worked out definition of a suitable demonstrator project in the time available, although attention was given to:-
The time might be ripe for the setting up of a Master's course on computer vision, and this possibility should be investigated.
Short courses on special topics would be of great value, both for postgraduates and for retraining established workers.
Object level languages specify robot behaviour in terms of how the objects in its world are to be moved around and brought into contact with each other.
Relative priorities of all these research areas:
Demonstrator project in robotics.
This will demonstrate an industrial robot being used in conjunction with other programmable devices to assemble a variety of objects, working from program which has been interactively produced and verified away from the robot. The emphasis will be on (a) the off-line programming system which should be easy for the ordinary person to understand and which makes use of sensory information, and (b) the run-time robot control system which should allow new equipment and new developments to be added to the system in a painless way.
We have divided the project into two parts, the programming part, and the control part. Each could stand by itself, but each would need the other for actual live demonstration. The robot programming language system is currently being developed with the support of the SERC Robotics Initiative.
Both of these are for a 2-3 year period.
Interested parties: Edinburgh University AI, Marconi Research.
Input to this system will be graphic and will take the form (a) of ketches input by a tablet, which serve to declare the most prominent features of bodies (b) the placement of secondary body features by specifying how they relate to primary features or each other, (c) specifying functional or geometric relations between features of different bodies Hearing B1 supports shaft S1.
Quantitative design data will be generated to satisfy the constraints input graphically, and a conventional Body Modeller will be used to display an object thus flashed out.
Interested Parties: Edinburgh University AI Dept, Marconi Research.
The use of an object level language for specifying high volume automatic assembly.
It is unlikely that multi degree of freedom arms can be developed to the extent of being able to perform several assembly operations per second. However, object level languages like RAPT can be used (with some development) to specify the behaviour of a new generation of machines for high speed manipulation which are more versatile than their predecessors. The manipulative elements of these machines will still only perform a single function, but this function will be made parametrically variable by the use of servo drives instead of earns. Object level programming will be needed to create machine settings and programs which are guaranteed to work correctly.
Post Script - Discussion on Transport.
Mr Popplestone said that the possibility of meeting the need to transport both goods and people within urban areas by the use of autonomous vehicles of moderate size running in segregated trucks should be examined. Mr Streeter, drawing on his experience of the abortive Sheffield mini-tram project said that the economics of such a system were problematic. It remains to be seen whether the increased versatility made possible by modern information technology and the change in attitudes to machinery arising from the expected dominance of information technology will change this state of affairs.
The development of IKBS should take user interfaces fully into account. It is important that they are not seen merely in terms of packaging but are considered as an integral and integrating component in systems. This is particularly important at demonstrator and task levels. The syndicate had 16 members, and the individual presentations themselves took place during the first 2 hours. We give an outline of these presentations in the next section before passing on, in section III to specific project recommendations.
Edmonds opened by considering possible approaches to the man-machine interface, and the role of cognitive studies. He emphasised that we should design the MMI from the outside, but looking in towards the heart of the system. He went on to outline current research projects at the Human Computer Interface Research Unit - Leicester. These include:- image handling, dialogue design, audio visual inspection, computer aided learning and expert systems in medicine.
Cowell presented ideas relating to his current research into computer aided learning systems which are implemented in COBOL and PASCAL on a PRIME computer. In the future it was hoped to explore intelligent interfaces for CAL, which would include research into appropriate languages and supporting editors, filing systems and HELP facilities.
Dr Bowlay presented a research proposal relating to the creation of Meta Data. This would enable the database to process certain knowledge of its own structure. The user would be able to ask questions such as:- What can I find out from this database? and How do I retrieve this particular information. In general he thought that the MMI should possess self-knowledge and also have some knowledge of user characteristics, which was particularly important with the proliferation of remote databases. In summary, he suggested that the aim of interfaces was to provide a window into the system and that IKBS was the appropriate research methodology.
Gray developed the theme of Meta Data and ways in which PROLOG might be used to allow databases to reason about themselves. In the longer term it would be expected that the complete database would be written out as a series of production rules. Indeed, it was possible that the interface processor could become larger than the pure database itself!
Eisenstadt opened with the comment that, Friendly systems are not in themselves good enough: It was necessary for the interfaces themselves to possess self-knowledge, and he proposed an assault upon four fronts:- i) AI, ii) Psychology/Cognitive Science - Mental Models, iii) Software Engineering, iv) Educational Psychology. As a possible demonstration project he proposed research into a Programmers' Aid.
Ewart discussed current research in his unit in which there was collaboration with both industry and the public sector. One current program was financed by the EEC for research into medical information systems. Ewart noted that the user contribution was vitally important in the design of successful interfaces.
Probert opened by emphasising that his perspective was that of the marketplace rather than that of pure research. He then discussed three possible areas for collaborative industry/university demonstration projects:- i) A "personal information assistant as part of an office automation package, ii) IKBS interfaces for corporate models and other decision support systems in BT, iii) Intelligent interfaces for Fault Diagnosis Equipment in the UK telecommunications network.
Fallside described his current research into speech training for the deaf and also into speech recognition. He then proceeded to outline two possible test-bed research programmes:- i) Speech operation and voice guidance systems for BT exchanges, ii) Speech in robotics. In this second project the aim would be to link spoken descriptions of a scene with a robotic vision system and so mesh speech production, understanding and visual analysis within the same framework. This project would thus be of multi-disciplinary interest, including links with linguistics and psychology.
Read summarised current research activities in the computing and sociology departments at Cardiff which related to research possibilities in MMI design. He focused upon database design and social data analysis.
Holmes divided his presentation between voice output and voice input, focusing first on voice output, he considered currently available equipment and ways in which improvements might be made to rule based systems through this IKBS research program. He then passed on to the nature of voice input technology in which he emphasised that a powerful voice I/O system will be an IKBS. Existing voice input systems were successful at reorganising isolated words, but only the most expensive of systems was able to reorganise speech in real-time to any acceptable degree of accuracy. He concluded by noting that IKBS and voice I/O need each other in all longer term research programmes.
Hopgood reviewed the state-of-the-art in MMI from several viewpoints including that of the psychologist. He noted in particular that the initiative should be carefully balanced between man and machine, and that machines must be adaptive to both the user and current context.
Kiss summarised current collaboration with industry within his unit which include joint projects with Unilever, ICL, Leyland Cars and ICL. He focussed his presentation upon the psychological aspects of the design of MMI and mentioned the MSc Course which is held at Warwick which includes the study of user interfaces from a psychological and cognitive perspective.
Harvey surveyed the interest of ITT in the domain of human factors and machine interfaces. He emphasised that we should develop interfaces which communicate at a range of levels according to the current skill of the user. The interface itself should be rich with many communication modes including voice, touch and graphics, keyboard, pointing and eye movements. He concluded by suggesting a demonstration project to develop an information sieve for electronic mail systems. This would filter out information which was relevant to the user from his/her electronic mailbox.
Self discussed his interests in Intelligent Teaching Systems (ITS). He was interested in the use of information retrieval systems by children using a natural language interface, when the children were unlikely to possess any keyboard skills.
Price surveyed various criteria used in the design of IKBS including physical, responsibility and contest criteria. This included the issue of whether we could trust the system! He also emphasised the difference between advisory and mandatory outputs from an IKBS.
Coombs opened by suggesting that research be initiated into the efficient modelling of the deep knowledge of user structure, and levels of understanding. He then went on to consider research at Liverpool into user interface for consultative systems and outlined a project entitled MINDPAD which was now being developed. He finalised by noting that research was also under way in collaboration with Unilever (through a CASE award), and also on an interactive interface language under a project called CONNECT. These last two projects were both under the supervision of Professor Alty, Liverpool University.
Courses are currently held at:-
Two current research contracts with BT (F Fallside - Cambridge, S J Young - UMIST) are concerned with use of synthetic speech and speech recognition (respectively) in computer controlled telephone networks. On completion in early 1984 these will provide useful practical and feasibility results and would provide a starting point for a demonstrator project. Already a number of developments are obvious, eg the use of synthesis from text and the use of natural language focusing for message generation in response to a particular request and as a result of particular network conditions. It is believed that BT would be interested in this project which would demonstrate the state-of-the-art in this area.
The system would enable users to input, edit and forward text, as in good conventional word processors. In addition it would allow images to be created or scanned in from TV cameras and for those images to be modified and merged with User text as required. As work at Lancaster Polytechnic has already demonstrated, user requirements for the manipulation of images in such contexts requires an intelligent system capability, in the context of a user-machine dialogue, of identifying elements in the images perceived by the user.
Interested: Edmonds and Scrivener, Leicester Polytechnic.
Timescale: 2 years.
A user asking for information through a teletext system (Prestel) would be able to describe views over the data. This involves both (a) presentation, layout etc, and (b) combining data from different sources, selecting it, transforming it and accessing other relevant facts based on content and association with initially retrieved facts. This could be used not just on Prestel pages but also on remote databases (possibly in combination). The view description must be based on access to metadata and the operation of describing the view should use combined visual, tactile and auditory inputs following MMI principles.
This would be a five-year project, aiming to develop a demonstration IKBS to provide automated debugging assistance for trainee programmers. The programmers would be working on a class of problems known in advance to the Programmer's Aid (PA). The PA would be able to detect errors in syntactically correct code, point out the cause of incorrect results, and (at the discretion of the instructor) suggest patches to the Trainee's code.
The class of problems to be tackled by the trainees would be at the level of difficulty of an elementary employee-payroll-update program, involving the following concepts: function, procedures, variables, argument passing, recorder, arrays, strings, conditionals, iteration, and a file I/O.
It is important to note that a software engineering approach based on current program verification techniques would (a) require the programmer to enter a detailed set of assertions before verifying the correctness of the program, and (b) would only be able to say whether the program was correct or incorrect (but not why). In contrast to this, an IKBS approach enables the programmer to enter code without bothering about assertions, and offers the prospect of reasoning about why a program has gone wrong. The key to this ability lies in the PA (a) associating segments of code with hypothesized intentions on the part of the programmer, (b) having a pre-stored repertoire of programming cliches and bugs observed in a large sample of trainees/programmes, and (c) having idealised effect-descriptions (in advance) for the class of programs which the trainee is trying to write.
The proposed advances in the state-of-the-art of programming aids are as follows:
Existing company models generally have user interfaces which are far from friendly in nature. Even with the addition of colour graphics and menu pad we find that the user still has to be very experienced in their operation. The aim of developing an IKBS interface is to enable managers themselves to use such models and decision support systems, in which the model understands its own operation.
The interface will essentially contain the expertise of an experienced user packaged as a series of production rules written in a language such as PROLOG. In addition, the model could in some sense analyse its own output and make recommendations to the manager as to policy options to be used for subsequent simulations.
Interested Company: British Telecom.
This is a suitable demonstrator project for IKBS and presents a particular challenge for man-machine interface design.
The marketable properties comprise:
Systems can be built now but improved man-machine interface (vision, pictorial, speech) essential to enable systems to penetrate longer areas of use (confined spaces, remote areas).
Systems will be required which are high integrity in the sense of transparently correct and free from hidden defects. This will demand transparency in the sense of metadata, distinction between evidence, hypothesis and inference. Systems must also be adaptive in the sense of matching their level of behaviour to the skill of the operator in order always to reach a conclusion in rapid time with minimum error.
Interested Companies: British Telecom and Marconi Avionics.
The scenario is that a user requires to access an information base such as a bibliography of literature covering a wide range of subject material. In accessing this base it is suggested that the user has inadequate understanding of both the structure of the information in the base and of exactly the nature of the information being sought. Keyword, author, and subject indexes are inadequate to meet this need since they depend on the indexers skill and on the users knowledge for their success.
A man-machine interface demonstrator project is proposed which will use a dialogue technique to establish a model of the users requirements and map these onto an IKBS model of the information in the system.
Interested Company: ITT Europe.
To provide an adaptable, general-purpose intermediary which can be used as an aid in driving complex but rule-based information systems like databases, operating systems, editors, application program packages, etc. It should be based on a powerful work-station type computer or microcomputer.
Interested parties: Warwick University, Psychology Department; ICI; Unilever; British Leyland Cars; BL Systems.
There is a rapidly growing interest in visual databases. Useful systems will require IKBS capabilities. In particular, the user interface will be difficult to build. Research is needed into such issues as methods of specifying image features to be used in a database search.
Interested parties: Edmonds et al, Leicester Polytechnic.
Timescale: Interim results in 5 years.
A person dialling in to a remote database needs to understand the nature of the data it contains. Currently this information about the data is usually stored in the tables of a data dictionary where it is relatively inaccessible to the beginner or casual user. Furthermore data dictionaries are limited to tables of relation and attribute names and do not, for example, hold information about constraints among the data or the meaning of the data in the database. We use the term Metadata for the much richer information about a database which is needed in order to properly understand the database itself.
The aim of the project is to build a user interface which enables a new user to explore the metadata of a wide variety of different knowledge bases/ databases by asking questions in a dialect close to natural language. The Advisor would answer users' queries about a database by accessing Metadata stored in association with that database. This will involve research into representations for metadata, and into methods of efficiently managing very large collections of Prolog clauses.
The user of a scientific database may well use a formal language such as SQL or relational algebra to extract and manipulate the rigidly organised data. But in order to understand things about that data he or she must interrogate the metadata, and the variety and free form of such queries make a natural language interface more suitable. For example, when querying a ship database the user may wonder why he or she gets no answer about the current ship position: the advisor can explain that the position readings are only taken weekly or that the time column refers to the time at which a message was received rather than to the time of taking the position and so on.
Important opportunities exist for the deployment and study of man-machine interaction techniques in the provision of facilities for authoring course material for teachers and trainers who are otherwise inexperienced users of computer systems.
IKBS systems are required at the man-machine interface to both support the author in writing his curriculum material and also within the authoring system itself to enhance user-workstation interaction.
Interested groups: Ulster Polytechnic; Cardiff University; Warwick University and Thames Polytechnic.
A test-bed project in the use of speech in robotics is proposed. In outline Robot 1 would provide spoken output describing a scene using a vision knowledge world and a speech production knowledge world. Robot 2 would be equipped with an automatic speech recogniser with an associated KW and would also have access to the same vision KW and speech production KW as Robot 1. This test-bed would provide useful results in several areas:
The project addresses fundamental questions in several areas and although proposed in a MMI syndicate, it is viewed as interdisciplinary. It has feasible short and medium milestones and in the long term would be extended to conversation between Robot 1 and Robot 2 with both equipped with vision, speech input and output and some form of inference.
Work at Leicester Polytechnic on the manipulation of instructed image words to be extended so that the user can, through an appropriate dialogue, describe to the computer what an image represents, or structure it according to his/her perception. Within the next few years further progress in this direction will require an IKBS of substantial power.
Interested parties: Edmonds et al.
Timescale: 5-10 years.
A message interpretation mechanism is seen to be a task generic to IKBS systems. The problem is exemplified in electronic mail systems where many messages exist of varying priorities and from sources which have some implicit social priority (eg a message from my boss will have a different implicit priority than a message from a colleague). An IKBS system managing this task would need to be adaptive to the users interpretation and definition of priority and message interpretation.
Interested Company: ITT Europe.
Under the broad heading of help facilities and debugging tools for the computer users of the future (including IKBS designers and implementers themselves), we see an important task for MMI workers to be developing intelligent programmer's aids. These aids should be more than (non-IKBS) software engineering tools, in that they should be able to reason about why a particular program has gone wrong, and to provide advice for correcting the source code. A demonstrator project based on this idea is described elsewhere in this report.
There is a body of psychological knowledge giving rules or heuristics for good man-machine interface design.
It is suggested that an MMI Design system be set up containing an IKBS encapsulating the psychological information and also information concerning rules for different types of users - novice, expert etc.
The system would query the application programmer concerning his requirements and produce an MMI which was effective in the environment and was acceptable to the user population.
See B5.
It is widely felt that user interfaces should be able to adapt to user words. Research is needed into mechanisms to effect this and into strategies for making such interfaces useful and beneficial. Much importance is attached to this issue.
Interested parties: Edmonds, Leicester Polytechnic; Coombs, Strathclyde; Kiss, Warwick; Hopgood, RAL.
The most important and urgent speech task in MMI for IKBS systems is to study how to use the latest existing speech I/O knowledge in such systems. An essential ancillary to this task is to develop methods of evaluating such systems, in terms of how the speech I/O actually helps the IKBS tasks, rather than in some arbitrary numerical terms about words correctly interpreted etc. The obvious way is to incorporate voice I/O in any appropriate demonstrator projects that are being done for their own merits.
The speech techniques chosen for this purpose should not be constrained to those of cheaply available commercial equipment, because such cheap systems are very limited in performance. Mere application of current technological processes could make existing research systems of low cost within a very few years. The necessary development investment is needed to get this done. However, even with the latest research successes, one should recognise that current performance, particularly on input, is far below what naive users might expect, and so the tasks should be chosen to take this fact into account.
I do not think at this stage that research on voice I/O techniques per se are correctly part of the IKBS programme, though there is a tremendous amount of voice I/O research that needs doing, and would properly form part of the Alvey programme.
Within a few years speech recognition will have reached the stage where IKBS techniques must be incorporated into the speech recognition system itself, but it seems a little premature to start this now. However, a study report on the way IKBS will be needed in speech recognition could well be undertaken.
See B2.
See C3.
The following are suggested for inclusion under the heading of capabilities as research topics:
Specification notations and implementation tools for dialogues have been developed for single stream input, relatively simpler systems. In IKBS now notations and tools will be required. Dialogue engineering will itself require an IKBS capability. Issues will include multi-stream input/output, enabling adaption and the monitoring of interface performance. This work will be important to much of the SPP.
Interested parties: Edmonds, Leicester Polytechnic; Hopgood, RAL; Alty, Strathclyde.
Timescale: 3-5 years.
See D1.
A sandwich design requiring two consecutive summer terms of taught courses including:
Students are expected to have computing degrees or equivalent qualifications/ experience. Leicester Polytechnic uses the taught posts as components in PhD programmes. PhD students from other institutions could attend them on the same basis given appropriate SERC support.
Among recent influences on the direction of psychology, the growth of the computer and allied sciences ranks very high. Workers in Cognition, Linguistics Control Theory, Information Science, Robotics and many branches of Psychology such as Cognition, Perception and Learning are coming to recognise the cross-disciplinary nature of their subject and increasingly useful interactions are anticipated.
This M.Sc. intends to reflect and encourage this new way to do Cognitive Science. The intention is to provide graduates in Psychology and other relevant subjects (see below) with a framework based on the interdisciplinary synthesis noted above, namely the acquisition, manipulation and use of knowledge. This will allow both theoretical and practical problems to be considered from the same viewpoint the aim being to keep the application of discoveries, techniques and theories in Cognitive Science very much in mind. Components of the course include:
Teaching will be by lectures and seminars, assessment by essays, examinations and a short dissertation in the form of a thesis, empirical investigation or programming project. The duration of the course will be 12 months.
The course is open primarily to graduates in psychology. However, other graduates, for example, in computer science, mathematics, linguistics, education, management science, library science, engineering, medicine, or some other subjects could also be considered, provided that the relevance of their interests to the objectives of this course is clearly established and that they know enough basic psychology. In appropriate cases the student may be required to follow a preliminary year of studies in psychology before embarking on this M.Sc. course.
The course does not form an integral part of the Department's Ph.D. programme, but it is envisaged that some of its graduates may be able to proceed to such a degree.
Graduates of this course could find employment in many areas of the rapidly developing technology of information handling systems, particularly those where problems of cognitive ergonomics are already apparent and can be expected to increase with the deployment of micro-processor systems into many new applications.
Apart from academic research and teaching, graduates of this M.Sc. course could be employed in the computer industry, software engineering, management, medical and legal information system design, business equipment manufacture (e.g. word processing systems), library information retrieval system design, educational technology, and other branches of information technology.
For further information contact: University of Warwick, Department of Psychology, Coventry CV4 7AL.
The MMI area needs to draw on psychological/cognitive science research whose support in the UK is insufficient at present.
It is suggested that the support of psychological research in the MMI/IKBS area is unambiguously adopted by the SERC on an adequate scale.
See F1.
Before summarising our discussion in terms of the 4 group topics, we wish to present what we see as the most important points to emerge:
Training and manpower. The shortage of trained manpower is a serious constraint on progress. Courses at all levels are needed, especially one year MSc. Some of these should be aimed at computer scientists, but there is a particular need to grain people with first degrees in applications disciplines who will return to these disciplines to develop expert systems applicable to them.
Traditional full-time courses of one year or more duration do not meet many of industry's needs. Alternatives are needed; both short term courses and sandwich type courses.
All courses require extra, earmarked teaching staff, equipment and support staff. Support staff are critical in a discipline which depends on sophisticated equipment. Without it, scarce teaching resources will be dissipated and expensive equipment underutilised.
To supply the teachers, fellowships for in service training are also needed.
Tools and Infrastructure. This is a difficult area, and the best languages and development environments are essential if we are to compete in the world market. The best of existing techniques must be made more widely available and a major effort is needed to extend the environment of existing languages. Beyond this, progress on better languages is essential to eventual success.
Communication must be one of the keys to building a successful research community. Networks and electronic mail are desperately needed. Support must be given not just to the control facilities but to providing access to peripheral groups, otherwise they will only be available to the large groups who may actually need them least. If we are to coordinate development with industry, industrial groups must also have access to these facilities.
Support for travel conferences and workshops is critical if workers are not to become isolated.
A national applications centre is required to act as a centre for resources and information. A journal and a good reference and bibliography service is badly needed. Particular effort needs to go into making technical reports from North America (and perhaps Japan) more readily available.
We turn now to an outline of the discussion, arranged by group topics:
We wish to stress the relevance of this field to IKBS.
eg a code of conduct for building and introducing an ES.
Almost everything in Section A can be viewed as being relevant here. We mention just the two most important dependencies on other areas:
As a prelude to our comments on the needs for Expert Systems Research, we note that the concept of an Expert System is not well defined, but the definition attempt at this stage is of little value. Nonetheless it was agreed that the concept extends well beyond the early models of MYCIN and PROSPECTOR.
Considering Expert Systems as a field of science, there is much to be done. There is:
Concerning the SPP Document itself, it was held that the list of demonstrator programmes needed revision. From the industrial standpoint, the choice of examples appeared uninformed. Overall, the research costs for the proposed SERC programme were considered to be too low and in need of revision.
First it must be very clear what the goals of such a programme are. We identified two objectives:
Despite the differing objectives, we consider that the state of the ES art luckily allows agreement on projects, if those projects can both improve academical understanding of questions bearing on basic principles of one subject, generate practical pay-offs which are of commercial areas which are ripe for such projects.
The industrial demonstrator exercise poses difficulties. We believe
However - the means of staffing this interchange are inimical to both University and Industrial management/career styles, and the resolution of this is not clear.
Specific difficulties include:
It was conceded that the academic reservations could perhaps be alleviated by paying them consultancy fees, and by freeing them from some of their administrative duties.
Within the syndicate a neat answer to the organisation problem was not forthcoming, however there was general agreement that some sort of small intermediate institution, staffed by both industry and university personnel, was the best prospect.
Specific examples of industrial demonstrator project should be determined by mutual negotiation between industry/university groups. Such examples can be foreseen in the fields of
but this general area list is rather mandated, nor complete.
No strong views were expressed, other than an encouraging pat on the back for parallel machines development, to the note that perhaps in this field recruitment from the USA is possible.
Studies are needed of the problem of inferring rules in an environment where the duration is subject to noise. The delivery of the inference rate probably exacerbate this sensitivity.
SERC should consider undertaking a large information gathering activity on inferencing applications with a view to producing a (Which of Inference Mechanisms).
Then, perhaps we can infer which ES technique suits our problem!
Above all the syndicate wished to stress the importance of non von Neumann languages and architectures, not just to the IKBS programme but to the whole of Computer Science. That these languages and machines are at the core of a revolution in Computing has been recognised by the Japanese in their Fifth Generation programme. The overwhelming majority of the syndicate were in sympathy with these conclusions.
Although the Japanese plans have triggered much activity in the U.K., particularly the Alvey Report and the IKBS SPP, it now seems clear that no such closely focused programme will be undertaken here. The syndicate, therefore. thought it necessary to re-emphasise the importance of a technology that is receiving much attention in several Computing and Communications sponsored programmes, particularly the Software Technology and Distributed Computing Systems programmes.
The syndicate therefore recommended that the main research themes to be pursued in this area were the continuing development of these languages and their associated machines and programming environments. This is not of course to say that more conventional languages and machines will not continue to be important and worthy of study and development but it was thought, particularly in the IKBS context, that the study of languages and machines that are able naturally to exploit parallelism was the paramount need. The research themes and topics outlined below are not meant to be in any way exclusive.
As important as the actual language used are the tools and disciplines provided to assist program development and maintenance in that language. It was felt that the declarative languages offer great opportunities for progress in these areas and work should be pursued to capitalise on this.
The syndicate thought it helpful to detail several projects it considered vital to progress in this area. Again no exclusivity is intended.
The syndicate identified a number of infrastructure or organisational requirements that need to be satisfied if such a programme is to be successful. Again this list is not meant to be exhaustive.
Unfortunately few experienced users of the kinds of systems we were meant to be discussing came to the syndicate. preferring to attend meetings discussing their area of research interest. We also felt under pressure to consider the ICL PERQ as more than just a possible machine because of the SERC's apparent (and, we believe, premature and irresponsible) commitment to it. Finally, we suffered from the fact that other closely-related common bases have apparently been decided; we did not have full details of them, and felt that earlier discussions on a common common base, as it came to be known, would have been sensible.
A common base is a set of facilities and software and expertise which is chosen and maintained or developed for the community at the community's expense with the results available free to the community. The overall cost effectiveness of the research programme is then improved because of the sharing of the cost of development, and also because of the portability (and thus sharing) of the resulting software products.
We would have liked to make the common base software-only, and allow a free choice of hardware, but the current state of the art in program portability, device-independent software and interface standardisation does not make this feasible. We have thus recommended suitable hardware too.
The free availability of the common-base software will make the choice of the recommended common-base hardware a cost-effective one for many projects, but it must be emphasised that the common base is not intended as a standard or norm. There will clearly be justifiable demand for both software and hardware that is outside the common base.
It is also clear that the common base must evolve, and that perhaps the most vital part of it is the mechanism by which this evolution takes place, Adequate funding must be available for the evaluation of new hardware and software and the commissioning of development efforts on behalf of the community.
The maintenance and distribution of common-base software (which will obviously evolve much more rapidly than the hardware) should be distributed. Each piece of software will be maintained at a site with appropriate expertise and a user community for that software; we see no place for a centralized software facility not engaged in active IKBS research.
At the moment the system is hierarchical and conventional, containing:
Actually the implementation-language is not an intermediate level, since it is a tool - the intended meaning is that it is the recommended implementation language for language systems and tools, in order to (a) ease system development and (b) improve the portability of the resulting software.
It is expected that in the near future language systems take over the functions currently implemented by (distinct) operating systems, such as process management and intercommunication, possibly across machines and current operating systems do not provide a suitable framework for this style. Research and development work in this area is encouraged.
Our overriding concern was that the systems chosen should be available immediately, with essentially no hardware or software development. Hardware should be available in the UK, with reliable maintenance contracts. Of course we also wanted systems that were already in use in the wider IKBS/AI community, in order to be able to make use of, and contribute to, other work. Our choice was:
This is currently not the cheapest way to support small systems, no the best way to support large ones, but it is available immediately and will be suitable for a large amount of the work That is required in the next few years. The most obvious deficiency is in the provision of high-resolution (1000 by 800) bitmap displays, and this area is one that should be investigated immediately, both theoretically (what is possible within the VAX architecture) and practically (what is actually available or coming soon).
Once the decision-making infrastructure is set up then further packages and tools will be candidates for the common base. Some areas have fairly clear needs and possible solutions (geometric modelling, hardware simulation) but others are more controversial (database systems) or problematic (dictionaries, images).
We felt that this term was essentially a political one, indicating a commitment on the part of the supplier to give large computing power to a user by making the sharing of his resources physically impossible, or at least difficult. Two features appear immediately - that a powerful display is possible because computer power is available and the user is local to it, and that inter-machine communication is vital. Currently there is no machine of this type that is suitable for effective IKBS research, reasonably low in cost, and supported in the UK. As a matter of highest priority this area should be surveyed, and community kept informed of developments so that due consideration can be given to extending the common base.
This machine does not have the right software available, and so cannot be included in the initial common base. Deficiencies in the current hardware, and uncertainty about future developments in it, mean that, while it should of course be included in the survey mentioned above, no IKBS funds should be spent on PERQ software development.
We felt that this topic came within our remit as there are technical decisions and standards required for it, though of course the mode of use of any network is in some sense an infrastructure issue. The highest priority is for mail, and then for file transfer. The ability yo work remotely is of much less importance, as links of the required bandwidth are probably too expensive. The Berkeley Unix system contains simple but adequate networking software which uses the public dial-up network; every machine must be provided with an auto-answer and auto-dialling modem until it is connected to some other appropriate and reliable network. gateways from this dial-up network must be provided to PSS and Arpanet. As a matter of urgency the community should mount an investigation of, and input to, current plans for networking in the UK academic and scientific communities. Note that the plans for software maintenance (above) and for networking imply that no central hardware facility will be needed.
Clearly work in the Common Base area should be coordinated with that in other SERC/DOI programmes. especially those on Software Engineering and Man-Machine Interface.
Courses involving the use of the kinds of systems we need for IKBS must form part of the education of many undergraduates, and this has considerable implications for the Computer Board, both in the type and the amount of resources they provide for University teaching.
Collaboration with industry will be vital in the development of a viable programme of IKBS research and development, so the tools and communications mechanisms mentioned above must be freely available to them. More than this, however, we believe that there are many industrial projects, both current and potential, which would benefit from access to current IKBS-type programming technology. This makes important the provision of relevant training (or other transfer of expertise), and makes urgent the investigation into future low-cost hardware for the common base.
