DISCOVERY AS COLLABORATION
It Takes Two (at Least) to Tango
Donald Michie
University of Edinburgh, UK
In his Structure of Scientific Revolutions,= Thomas Kuhn distinguishes between:
(1) "discoveries, or novelties of fact" and
(2) "inventions, or novelties of theory".
Examples that have been cited of the first are the= discovery of the New World by Columbus, the discovery of the planet= Neptune, the discovery of new chemical elements and the discovery of= penicillin.
Examples of the second category are Newton's theory of= gravitation, Darwin's theory of natural selection, the microbial theory of= infectious disease and the greenhouse theory of global warming.
Traditional thought holds that theory discovery= typically proceeds in discontinuous "insights of reason" (Aristotle),= "creative intuitions" (Bergson), "inspirations" (Popper), "flashes of= genius" (Reichenbach) and the like. These "interventions from= nowhere" lie beyond immediate introspection. They are commonly= ascribed to intrinsically mysterious and possibly unknowable causes.
A similar position could be (but is not) taken= about fact discovery. Householders who have forgotten where they have left= something not uncommonly testify that their wanderings took them "for= no particular reason" to the approximate whereabouts of the lost= object. When discovery immediately follows, protagonist and onlookers do= not ordinarily refer to a "hidden hand" or to a "divine= spark". Similar restraint moderates the pride of fact-discoverers of= greater moment. The following is early Portuguese testimony to the theme of= background knowledge to which we shall later return:
They discovered new islands, new lands, new= seas, new peoples, and what is more important, new heavens and new stars = =85 Now it is clear that these discoveries =85 were not achieved through= guesswork: our seamen set off well trained and provided with instruments= and rules of astronomy and geometry.
from Pedro Nunes, 1537.
Person and Zombie
Household fact-discovery is sufficiently mundane= that there is no threat to human dignity in supposing that perceptions and= inferences come from below the level of the conscious Person. What= neuroscientists tend to refer to as "the Zombie" probably guided= the Person=92s steps, as the only means of signalling. But in the grander= arena of theory-discovery, the mind=92s sudden insights are more often= presented as coming from nowhere.
The thrust of this paper is that software simulation of= theory-discovery, just as much as of fact-discovery, depends on abandoning= the model of a unitary mind visited by flashes. Modern brain science views= the overwhelming bulk of human perceptions and behaviours, including those= of abstract thought, as driven in fast time by hidden zombies. The Person= reserves awareness for a few strategically critical functions of planning,= monitoring, rationalization and communication.
The singular noun "Zombie", like= "Army", here denotes myriads of unseen agents, mutually= co-ordinated. As in Tolstoy=92s portrayal of Napoleon=92s role in battle,= the Person=92s reward for being swept along by the stream of events is the= privilege of issuing commands and requests for information, and of= exchanging messages with allied generals (other Persons). Tolstoy=92s= contemporary, the British scientist Francis Galton (1883), was already= describing the conscious mind in similar terms:
The position of consciousness appears to be= that of a helpless spectator of but a minute fraction of automatic= brainwork.
In the time-scales of professional= tennis or lightning chess Galton=92s account is today known not to be an= exaggeration. In slower time, though, the Person=92s role is a good deal= more than that of mere commentator. Pro-active collaboration by the Person= with the hidden zombies then emerges, and low-capacity channels between the= two open up. The following specimens of signalling between the two are= consistent with modern knowledge.
From Person to Zombie (used plurally, as in= Army)
From Zombie to Person
The cello-playing example is from K.= Furukawa=92s discussion elsewhere in this Volume.
For cognitive simulation, opacities found in= nature must be preserved. But from the Person=92s point of view, the= Zombie=92s opacities can be inconvenient. For the technological purposes of= machine intelligence we can aim higher. Experimental demonstration= of the extraction of transparent "behavioural profiles" from a= fast real-time skill was shown by Michie, Bain and Hayes-Michie (1990; see= also Michie, 1998). In effect, it is possible to populate even the lowest= levels of fast-time problem-solvers with "super-articulate= zombies." A paradigmatic case study from A. D. Shapiro (1987) will= later be presented.
The phenomenon of "blind-sight" will now be= described.
A case of "blindsight"
The following clinical condition is illustrative of= Person-Zombie duality. When the primary visual cortex has been compromised= by stroke, trauma or surgery, the patient is partially or wholly blind. But= the phylogenetically more primitive "orienting" visual pathway is= still intact. Even though the Person is unaware of what is going on, this= pathway can capture and utilize visual inputs, thereby generating tell-tale= behaviours in the seemingly blind. Here is a patient of Sanders and= Weiskrantz, referred to in Ramachandran=92s (1998) Phantoms in the= Brain as Drew.
They held up a stick, in either a vertical or= a horizontal position, in his blind field and asked him to guess which way= the stick was oriented. Drew had no problem with this task, although he= said that he could not see the stick. After one such long series of= "guesses", when he made virtually no errors, he was asked,= "Do you know how well you have done?"
"No," he replied, "I didn=92t =96= because I couldn=92t see anything; I couldn=92t see a darn= thing."
Note that the problem posed by the= experimenter=92s guessing game was necessarily solved by a collaboration= between two conceptually and experimentally separable agents of the= patient=92s brain. The Zombie made the stick-orientation discoveries. Drew= (the Person) somehow unknowingly glimpsed them, before retransmitting them= as guesses.
Duality Principle
Muggleton and Michie (1996) have proposed a= somewhat similar two-layer software architecture for AI, referring to it as= "The Duality Principle":
=85 software involved in human/computer= interaction should contain two distinct layers =96 a declarative= knowledge-level layer and a lower-level functional or procedural-knowledge= layer. This extends the formal methods separation of specification and= implementation by requiring that the declarative layer be capable of= extensive human interrogation at run time.
An advantage of the artificial over= the natural system is that the knowledge-level Person can be endowed with= better communication with run-time zombies than is enjoyed by the conscious= brain. In the case of the forgetful householder, a common pattern is for= the person to review in declarative memory what can be recalled of the= circumstances surrounding the last sighting of the lost object. Having= found it, with or without subconscious promptings, in the mental model= ("Aha!"), the person then launches appropriate zombies to support= real-time exploration of the newly recalled vicinity. In intelligent= software, the two-layer operating systems of the future will doubtless= short-circuit such mechanisms. Direct inspection of traces not only of= top-level transactions but (critically) of histories and states of= subordinate agent-modules could become routine.
One of our agents is missing
The foregoing stick-guessing example illustrates= the brain=92s duality principle from the perceptual side. Below is an= action-oriented case chosen to illustrate a form of fact discovery usually= known as problem-solving. Again it comes from Ramachandran (1998). The= problem is simple arithmetic. The patient had suffered a small localized= stroke affecting the brain=92s left angular gyrus, already known to be= specifically associated with numerical calculation.
At one point I said, "Okay, Bill, can= you subtract seven from one hundred? What=92s one hundred minus= seven?"
He said, "Oh, one hundred minus seven?"
"Yeah."
"Hmmm, one hundred minus seven."
"Yes, one hundred minus seven."
"So," said Bill. "One hundred. You want= me to take away seven from one hundred. One hundred minus seven."
"Yes."
"Ninety-six?"
"No."
"Oh," he said.
"Let=92s try something else. What=92s seventeen= minus three?"
"Seventeen minus three? You know, I am not very= good at this kind of thing," said Bill.
"Bill," I said, "is the answer going to= be a smaller number or a bigger number?"
"Oh, a smaller number," he said, showing that= he knew what subtraction is.
"Okay, so what=92s seventeen minus= three?"
"Is it twelve?" he said at last.
Further questioning revealed that= Bill had an intact and quite sophisticated model of the number system,= including infinity, approximation, relative magnitudes and the rest. Only= the calculational agent was missing from the library of procedures, leaving= the library manager (Bill) without resource whenever this specific function= was tested.
Imagine, then, a Turing Test confined to the domain of= simple arithmetic. The respective stances of the logicist school ("let= an artificial Bill work it out from first principles") and of the= reactivist school ("leave everything to the zombies") both fail.=
Child Machine
Now unite Person and Zombie in a partnership, with= one additional responsibility for the Person, namely= zombie-training. Result: a three-layer prescription for= skill-acquisition reminiscent of Turing=92s (1950) "child= machine". Turing asked why we should not break the task of developing= a machine intelligence into two steps, essentially:
A modern up-date might further= refine this:
Layer 1 must incrementally take over= some of the human teacher=92s tasks, acquiring further facts and= generalizing theories for itself. Layer 1 must also take a hand in the= building of layer 2, i.e. in zombie-training. By this we mean that it must= acquire certain declarative elements of the human trainer=92s craft= concerned with setting the framework of task-segmentation, subgoals and= constraints within which level-2 re-inforcement learning and classification= learning proceeds.
Returning to Bill, long before his calculating zombie= was disabled by the stroke, he had early lost conscious touch with it - as= we all do once arithmetic skills have become automatic. Yet his youthful= Person, goaded and guided by teachers, had personally trained this zombie= in the first place, by setting it goals, rehearsing it in endless practice,= and using declarative-layer knowledge to correct errors.
News from nowhere
So what of novelties of theory rather than= novelties of mere fact? Testimony of creative thinkers, from the novelist= Henry James to the mathematician Henri Poincare is in line with the idea= that flashes of insight (illuminating the declarative layer) are preceded= by systematic and repetitive thought-experiments accomplished by largely= unconscious brainwork in the functional layer:
For a fortnight, I struggled to prove that no= function analogous to those I have since called Fuchsian could exist; I was= then very ignorant. Every day I sat down at my work table where I spent an= hour or two; I tried a great number of combinations and arrived at no= result. One evening, contrary to my custom, I took black coffee; I could= not go to sleep, ideas swarmed in clouds; I sensed them clashing until, as= it were, a pair would hook together to form a stable combination. By= morning I had established the existence of a class of Fuchsian functions = =85 I had only to write up the results, which took me a few hours.
Henri Poincare, quoted by Bell (1937)
The results, but often not the= workings themselves, then leak into conscious awareness. They are commonly= perceived by the discoverers as creative flashes. On this view such= insights come "from nowhere" in the same sense that the= blindsight patient=92s guesses come "from nowhere". Evidently it= is necessary for us humans to be able both to sense phenomena and to attend= to the world to which they relate. Only then do we have a= "somewhere" to put them.
Ramachandran=92s patient Ellen is unable to attend to= anything in her left visual field. She contrasts with Drew in that if her= attention is drawn to a left-field object, her intact primary visual cortex= allows her to see it. But damage from a stroke to Ellen=92s right parietal= lobe has caused the condition known as "hemineglect." The stored= mental map of her world lacks a left half. It is interesting to see what= happens when this visualizable world fails to support her logical world. In= the latter world, clock faces must have 12 inscribed numbers, as entailed= by known facts about the times of day and night.
When I asked her to draw a clock, Ellen made= a full circle instead of just a half circle. This is a fairly common= response because circle drawing is a highly overlearned motor response and= the stroke did not compromise it. But when it came time for Ellen to fill= in the numbers, she stopped, stared hard at the circle and then proceeded= to write the numbers 1 to 12, cramped entirely on the right side of the= circle!
V.S. Ramachandran and S. Blakeslee, 1998
Two case studies
The Duality Principle will be illustrated in two= contexts. In both, agents have performed feats of collaborative discovery= unthinkable without inter-agent collaboration. In the first case the= collaboration is between members of two castes of ant. In the second it is= a three-way collaboration between a chess-master, a programmer and a= developing Turing-style "child machine". Each discovery presents= features that outrun what an individual human could accomplish.
The first concerns discoveries of fact. The discovery= agents were members of various species of Siberian wood ants. Discovery by= the ant teams of new food sources are of course limited to "novelties= of fact". The only "inventions" are the collaborative= mechanisms themselves, designed and built in this case over tens of= millions of years of evolution.
In the second, the agents were human-computer complexes= and followed A.M. Turing=92s "child machine" prescription for= collaboration between an expert human teacher and a teachable program. The= application was to a chess end-game that had previously defeated all= attempts to construct a viable theory.
The respective claims made are that:
Both run strongly counter to the cultural flows of the= respective fields of ant biology and of AI. Seismic shifts of perception= and practice may be required if their plain message is to be assimilated= into either scientific canon.
Case study 1. Symbolic communication in ants
The following account is based on my report to the= Royal Society of London on a visit to Siberia during 18th June =96= 2nd July 1998, jointly supported by the Royal Society and by the= Russian Academy of Sciences.
The Siberian work conducted by Dr Zhanna Reznikova= forms a structured sequence based upon more than 15 years of systematic= field and laboratory studies. Her husband and co-worker, Dr Boris Ryabko,= is a mathematician specializing in coding theory. His role in the work has= been design, analysis and interpretation. In spite of forfeiting visibility= by scanty choice of Western publication outlets, both workers are= independently known and internationally respected for sustained studies in= their respective fields. Some information theoretic aspects of the work are= summarized by the authors in Complexity, 2 (2), 37-42, Wiley= 1996, ISSN 1076-2787 (see http:// journals. wiley. com/). References are= there given to some of their English-language experimental reports on ant= cognition and communication.
The experimental series began in the early 1980=92s= with Reznikova=92s refutation of a dictum still found in Hoelldobler and= Wilson=92s 1990 treatise The Ants, that ants cannot recognise each= other and thus cannot form and maintain specialized teams (=91cliques=92 in= their terminology). The three species used in these studies base their= economy on outdoor foraging work conducted by just such teams. In the= Siberian work, indoor colonies are maintained in the laboratory. Every= participating ant is individually marked. Each team consists of a scout,= who prospects for food hidden in experimentally constructed mazes, together= with her foragers who then go and get it. Between these two phases are= interposed the following events:
(1) return by the scout to the nest to locate and= mobilize her own team,
(2) a scout-to-foragers communication session using= antennae, legs and palps, lasting from tens of seconds to many minutes and=
(3) an excursion of the team (minus their scout who= has meanwhile been experimentally removed) to the food location.
Key move: record the durations
In the months before my Russian trip I studied the= authors=92 ten published English-language papers and questioned them= via email. Their key move was based on the fact that they knew on= each occasion the location of the hidden food, having themselves placed it= in a selected end-point of an experimental maze. In the default condition= selection of end-points was made at random. In one critical variant, random= selection was conditioned on a strong bias towards one, and in some cases= two, pre-selected end-points.
The experimenters were thus able to make inferences= about message structure from recorded message durations paired with= the known complexities of their content. The experimenters pre-determine= the content (i.e. what information the scout must transmit if her foragers= are to go to the right place) through their freedom to construct different= topologies of maze and different end-points in the maze to place the food.=
In a given experiment records of each of the= constituent trials have the form
Trial no., Date, Known location, Message= (secs), ID of team.
On her return to the waiting team,= there are two types of possible question that the scout=92s message might= be framed to answer, namely:
(1) =91What must we do to find the food?=92=
or alternatively
(2) =91Where is it?=92
The possibly subtle-seeming difference between (1) and= (2) has ramifications. It is a question of whether message recipients can= interpret only command-strings, or relational descriptions as well. This= question, although important, was deferred. In the experimental series on= which I concentrated, the scout has only to transmit to her team a single= integer, corresponding to the ordinal number of the side-turning in a= sequence of some 40, at the end of one of which the food lies. In this case= a sufficient explanation of observed behaviours can be constructed in terms= of (1) alone. Other experiments (mentioned but not reviewed here,= see =91Co-ordinate grid exams=92) yielded suggestive evidence that= the ants can indeed also form, transmit and use "cognitive maps".= In all categories of experiment I satisfied myself that laboratory= procedures had excluded obvious pit-falls, such as ant-to-ant communication= of directions via odour trails.
My two-week visit focussed on the results of= experiments designed to explore ant capabilities to encode, to decode, and= to change the coding base, of messages each of which essentially encodes a= single integer. I do not here include the =91superlong=92 messages, see= later.
Determining message durations entails only the use of a= stop-watch, and an objective criterion for recognising the beginnings and= ends of transmission. Mathematical cryptography, in particular Shannon=92s= statistical theory of information and Kolmogorov=92s algorithmic= information theory, allows much to be inferred about the means of encoding= just from message-lengths alone, - given that one already has a crib, in= the sense of having advance knowledge of the contents of messages. By the= time of my arrival on 18th June 1998 I had pared my concerns= down to a short list comprising a few pages of handwritten notes, and had= conjectured that two weeks on the spot should decisively settle two key= points concerning counting.
The first point had essentially already been= established by published material, namely =91Does the scout transmit to her= team integers in a unary code? =92 The answer has repeatedly been shown to= be =91Yes=92.
The more controversial claim was that, faced with a= radical change introduced by the experimenters in the relative frequencies= with which particular integers are associated with food, the scouts change= their numbering system so as to exploit the new regularities in the= interests of message-length economy. The latter is important since= transmission rates are of the order of only a couple of bits per= minute.
Superlong messages
I had set out on my journey in a mood to reject the= investigators=92 second claim. Their English-language publications= contained no hint of events that seemed to me inseparable from the= hypothesised change of number code, namely superlong communication sessions= intervening between the old-code phase and the new-code phase. Such= superlong sessions would surely be needed for each scout to indoctrinate= her team in new-code.
In prior email exchanges, Dr. Ryabko had argued against= the logical need to expect such tutorial sessions. He did, however, own= that, although unreported outside the Russian literature, such unexplained= long sessions had indeed been observed. During my stay, I was able with Dr= Reznikova=92s help to follow this up. We combed her original logs and found= internally consistent and mutually supportive instances of the postulated= superlong sessions. Preliminary analyses of the contexts in which the= superlong sessions occur, together with a related phenomenon of 'mass= mobilization' also recorded in the logs, convinced me that a prima= facie case exists for the =91new-code=92 claim. I was able to bring= back photocopies of the most relevant original documents for further= analysis.
The new discovery takes the study of animal= communication beyond the widely-studied but limited powers of symbolic= communication found in the honey bee. For one thing the ants=92 signalling= system embodies a digital as opposed to analogue code. It also calls for a= revision of ideas about the generation and transmission of cognitive= content, formerly thought to be restricted to higher vertebrates. It seems= therefore worth placing the specialized numerical capabilities which I= selected for my study visit within the broader experimental framework that= the Siberian workers have constructed over the years.
Experimental framework
With the three species that had been shown to have= scout-forager teams, all experiments have two parts. The run-up consisted= of familiarization trials which sometimes lasted as long as a month. During= this period a minority of scouts with poor records of competence are= identified and discarded. The run-up is followed by lengthy and rigorous= 'final exams'. The investigators claim to have established the following= propositions.
1. Binary tree exams
The scout can routinely impart to her team= =91left-right-left . . .=92 directions for routes encompassing up to six= forks. Message durations are linearly dependent (except as qualified below)= on the Shannon complexity of the message, - e.g. a path involving four= left-right choices rates 4 bits.
However, for fixed number of left-right choices,= different routes have different human-perceived complexities. Thus LLLLLL= (six 'left' turns in a row) should be more briefly expressible than LRLRLR,= which in turn should have a shorter encoding than the formless LRRRLL).= When message length is fixed, these preference-orderings correlate highly= and consistently with the corresponding ant message duration, and indicate= a Kolmogorov complexity component.
2. Number exams
In Formica polyctena most, but not all, scouts= can count and reliably transmit integers up to about 40.
The scout communicates the results to her foragers in a= unary code (e.g. a human user might phrase =914=92 as =91finger finger= finger finger=92).
She can re-optimise her numerical code in response to= experimental modification of the statistical properties of the task. The= modified codes use simple additions and subtractions, achieving economy in= a manner reminiscent of the Roman numeral system when compared with =91fing= er finger finger ...=92
3. Co-ordinate grid exams
A scout who finds food at one of the intersection= points of a rectangular grid of stout wires, transmits a message of a form= quite different from the =91binary tree=92 instruction sequences. Here the= transmitted information appears to convey the co-ordinates of the distant= location. The team goes to the spot, but each by her own, sometimes= roundabout, path. It was to these experiments that I referred when speaking= earlier of "cogntive maps". Although strongly suggestive, the= data are less massive than those of series 1 and 2 above, and lack= year-on-year repetitions.
The three species concerned are Formica= sanguinea, Formica polyctena and Camponotus saxatilis.= Informal trials of saxatilis in the numbers game were disappointing.= But other observations show that on some tasks they can be as smart as the= others. All-round, F. polyctena score best.
Points of perspective
Some points of perspective are needed. The= observations apply only to a few =91freak=92 species of ants, i.e. selected= from the ten or so that have been described as highly socialized out of a= total of some 6000 species. Today=92s Homo sapiens has experience of= a densely urban life-style which places a premium on symbolic= communication. This experience belongs only to recent, i.e.= post-evolutionary, history. Otherwise such lifestyles are confined to the= social insects, notably the ants, termites and bees. There is a difference.= Experience of complex social existence is known to reach back many tens of= millions of years, necessarily if required skills are to be evolved as= innate routines. Acquisition via genetic evolution is indeed the= only path. Their relatively simple nervous systems, coupled with the= impossibility in ants of inter-generational cultural transmission, rules= out the human trick of inventing and culturally transmitting symbolic= communication as required.
To elaborate on the role of nature versus nurture, the= numerical and communicative competences found in these species of ants must= have arisen in the genetic code over evolutionary time, rather than being= re-discovered in each generation by brains innately lacking built-in= symbolic skills. For these ants cultural inheritance is not an option. Each= new colony is founded afresh by a single fecund queen. Whatever is learned= in the previous generation is lost.
Anything resembling cognition in an individual ant has= to rest on a rather small neuro-anatomy. The part of an ant=92s brain (the= "mushroom" structure) that is deemed equivalent to the vertebrate= cerebral cortex contains a mere quarter of a million neurones. In the human= cortex the corresponding number is ten thousand million.
The other salient contrast in this instance between= human and ant problem-solving is in the caste-based factorization of a= total cognitive task into components. Thus:
Since the period of my visits Dr Reznikova has renewed= her attack with a graduate helper on relating the capabilities revealed in= the laboratory set-up to behaviours observed in the natural tree-top= working environment. Every night the ants return to mound-like nests at the= foot of the trees. The following is from an email message recently= received:
With Tanya Novgoroda we observed antennal= contacts by which scouts attracted aphid milkers to the new aphid colony.= The scout mobilizes the team from the neighbouring branches of the tree,= not from the nest (this would be too far). This suggests a use for symbolic= communication: to solve the very important problem of providing the ant= colony with honeydew gained from aphids =85 the ant colony depends on= honeydew to a great extent: they eat more than 1000 kg (in dry sugar= weight) per summer. And their success depends on success of scouts which= search for new aphid colonies in a very changeable situation (rain,= predators and so on).
Reznikova and Novgoroda also found= that having been mobilized by their scout to proceed to a designated leaf= to farm a new herd, the foraging team in Formica polyctena exhibits= further subdivisions of function related to their tasks of= aphid-management. Subtasks include mounting guard, herding and milking. It= is not yet known whether these roles are exchangeable within the forager= group.
Results of extensive tests suggest that foragers cannot= communicate symbolically with each other. Technically they are of a= different caste from scouts, who are somewhat smaller. In the wild, one= scout is responsible for the management of several teams.
Duality Principle again
The hierarchical scout/foraging teams arrangement= is not strictly to be compared with the Person-Zombie dichotomy in human= cognition. It none the less presents a layered structure of collaborating= agents. To that degree it offers suggestive parallels.
Advances are currently being made in continental and= British laboratories along several dimensions of ant cognition, including= the combination by desert ants of global and local frames in support of= navigation by landmarks. Separately from what was described above,= Reznikova can reproduce in ants cognitive communications of a kind that= raised a stir in the West when first elicited in dolphins (see Andersen= (ed.) The Biology of Marine Mammals, Academic Press, 1969). With= dolphins, communication was peer-to-peer. With Reznikova's ants it was= scout-to-forager.
Case study 2. Discovery of a complex theory in= chess
An overt version of Turing=92s proposed= child-machine method was applied in the early 1980=92s to a chess endgame= micro-world, namely King and Pawn (on a7) versus King and Rook= (conventionally written KPa7KR). The program was required to adjudicate,= giving its reasoned justifications, claims by the White player (who in this= class of positions threatens to promote the pawn to victory) that a given= adjourned position is won for White. The resulting hierarchical= concept-driven program was subsequently able to demonstrate its chess= intelligence by fielding, and commenting on, questions put to it at= an international gathering of computer chess specialists and chess-masters= (see Michie, 1986; Michie, 1995a).
In doing so, it immeasurably outran the primitive and= sketchy grasp of this end-game which expert human analysis had previously= achieved over a century or so of intermittent study. Note that what is here= under discussion is not competence in correctly attaching the label= "win/not-won for White" to arbitrarily presented positions: for= KPa7KR, Grandmasters can do this with near-100% accuracy. What is at issue= is an agent=92s ability to produce complete and reasoned justifications for= each act of labelling. For KPa7KR, chess-masters are very far indeed from= this.
A form of miniature Turing Test was thus found on this= occasion to contribute evidence of the method=92s adequacy for a specified= task, namely the inculcation into a computing system of what would be= classed as a highly developed form of understanding if manifested by a= human.
More than a decade has passed since this first educable= child machine was engineered by A. D. Shapiro (1987). He termed the= teaching-by-examples method "Structured Induction". The teacher= was International Master D. Kopec. Both worked at the Turing Institute,= Glasgow. A version of the educable child machine itself was then made= commercially available to an industrial client of the Institute.= Westinghouse Electric Corporation at Pittsburgh, USA, had been faced with a= seemingly intractable problem of optimising, not chess end-game strategy,= but automated quality control of a nuclear fuel refining plant.
For the end-game task, generations of end-game scholars= had failed to build even the rudiments of a theory. The industrial= problem=92s intellectual complexity proved to be considerably less. A staff= member (Leech, 1986) using the identical methodology was able in a few= months to teach a closely related related software package to act as a= high-performance automatic controller. The consequent marked improvement of= yield efficiency brought savings to the company exceeding $10 million per= year.
Since then the use of the child-machine software and= methodology has permeated a specialized sector of industry. A spot sample= of seven case-studies from a single firm is summarized in the Appendix A of= Michie (1995b). After Shapiro=92s monograph went out of print and he had= left for a career as a commercial programmer, academic interest lapsed.= Recently a partial form of structured induction was revived by an academic= group for a pharmacological industry problem. Muggleton, Bryant and= Srinivasan (2000) used Inductive Logic Programming (ILP) in something like= the same style for discovery of rare structure-activity relations (SARs) in= neuropeptides. The estimated reduction of commercial search costs was= 1000-fold. The ILP tool used, namely Muggleton=92s Progol, continues to= accumulate chemical knowledge coded in first-order logic for use in= computer-aided discovery of further molecular chemical Structure Activity= Relations.
Turing=92s 1950 child machine proposal envisaged= precisely such a path, involving what in machine learning is today termed= "background knowledge". Turing=92s prescription (Mind, p.= 457) was that the educable system should have "a complete system of= logical inference built in. =85 The store would be largely occupied with= definitions and propositions. The propositions would have various kinds of= status, e.g. well-established facts, conjectures, statements given by= authority, expressions having the logical form of propositions but not= belief value."
In the micro-world of biomolecular theory-discovery,= not only is the child machine approach scoring early successes but the= spirit of the Turing Test unexpectedly re-appears. The biochemical= "teachers" of these prototype systems report that they can only= effectively critique the machine-generated molecular conjectures if they= can credibly see these conjectures as emanating from a human colleague. To= facilitate this the Muggleton group have retrofitted a primitive= English-language front end. Turing-style imitation games for assessing the= biomolecular intelligence of candidate theory-discovery programs have a= foreseeable future place in the pharmacological industry.
Conclusion
The AI research community has in the past viewed= intelligence, whether applied to fact discovery or theory discovery, as a= unitary system. Just which of two behavioural paradigms is the key -= whether Person-like or Zombie-like - has been a topic of dispute. There is= an indication in this that a point has been missed. The key, if anywhere,= is not in either, but in the interaction. Development of the arts of= Zombie-simulation and Person-simulation should be seen as separate= preliminaries to integrated collaboration between notionally separable= components. The intention of my two case histories has been to illustrate= this point. There may be worse places for machine intelligence designers to= look than among the varieties of interaction between the brain=92s layered= agent systems and among other layered forms of inter-agent collaboration= found in nature.
Finally, it will not have escaped notice that the= material reviewed in this paper lends strong further support to Marvin= Minsky=92s (1986) broad philosophy of the human mind as a multi-agent= system.
References
Bell, E.T. (1937) Men of Mathematics.= London: Gollancz.
Galton, F. (1883) Inquiries into Human Faculty and= its Development. London: Macmillan
Hoelldobler, B. and Wilson, E.O. (1990) The= Ants, USA: Harvard; Berlin, Heidelberg: Springer-Verlag.
Michie, D. (1995a) Consciousness as an engineering= issue, Part 2. Journal of Consciousness Studies, 2 (1),= 52-66.
Michie, D. (1995b) Problem decomposition and the= learning of skills. In Machine Learning ECML-95, Lecture notes in= Artificial Intelligence, 912, 17-31. (ed. N. Lavrac and S. Wrobel),= Berlin, Heidelberg, New York: Springer Verlag).
Michie, D. (1998) Learning concepts from data.= Expert Systems with Applications, 15,193-204.
Michie, D., Bain, M. and Hayes-Michie, J.E. (1990)= Cognitive models from subcognitive skills. In Knowledge-based Systems in= Industrial Control (eds. M. Grimble, J. McGhee & P. Mowforth),= Stevenage, UK: Peter Peregrinus, pp. 71-99.
Minsky, M. (1986) The Society of Mind. New York:= Simon and Schuster.
Muggleton, S.H., Bryant, C.H. and Srinivasan, A. (2000)= Learning Chomsky-like grammars for biological sequence families, Proc.= 17th Internat. Conf. on Machine Learning, Stanford= Univ. June 30th.
Muggleton, S. and Michie, D. (1996) Machine= intelligibility and the duality principle. BT Technol J, 14= (4), 15-23.
Ramachandran, V.S. and Blakeslee, S. (1998) Phantoms= in the Brain, London: Fourth Estate Ltd. (paperback, 1999).
Ryabko, B. and Reznikova, Zh. (1996) Using Shannon= entropy and Kolmogorov complexity to study the communicative system and= cognitive capacities in ants. Complexity, 2 (2), 37-42, John= Wiley. Online ISSN: 1099-0526. Print ISSN: 1076-2787.
Shapiro, A.D. (1987) Structured Induction in Expert= Systems. Addison Wesley.
Turing, A.M. (1950) Computing machinery and= intelligence. Mind, 59 (236), 433-460. Reprinted 1992 in= Mechanical Intelligence: Collected Works of A.M. Turing (ed. D.C.= Ince), Amsterdam, London, New York, Tokyo: North-Holland.