0% found this document useful (0 votes)
554 views22 pages

Pandemonium Selfridge

Pandemonium: a model of a process which we claim can adaptively improve 1 tself to handle certain pattern recogrn t I on problems. The basic motIf behind our model is the mrttrm of parallel processing. It is often easier to handle data in a parallel manner, and, secondly, it is easj e r to modify an assembly of quasi-independent modules.

Uploaded by

Khem Caigan
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
554 views22 pages

Pandemonium Selfridge

Pandemonium: a model of a process which we claim can adaptively improve 1 tself to handle certain pattern recogrn t I on problems. The basic motIf behind our model is the mrttrm of parallel processing. It is often easier to handle data in a parallel manner, and, secondly, it is easj e r to modify an assembly of quasi-independent modules.

Uploaded by

Khem Caigan
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 22
NATIONAL PHYSICAL LABORATORY SYMPOSIUM No. 10 Mechanisation of Thought Processes Proceedings of a Symposium held at the National Physical Laboratory on 2gth, 25th, 26th and 27th November 1958 VOLUME I LONDON: HER MAJESTY’S STATIONERY OFFICE 1959 (Reprinted 1982) SESSION 3 PAPER 6 PANDEMONIUM: A PARADIGN FOR LEARNING by DR. 0. G, SELFRIDGE BIOGRAPHICAL NOTE Oliver G, Selfridge was born in London 10 May 1926, He studied et the Massachusetts Institute of Technology fron 1942-1948, returning post~ praduately from 1946-1950, After 2 years at Signal Corps Laboratories at Fort Monmouth, New Jersey, he jolned Lincoln Lavoratortes in Group 84, Conmunteation Techniques, of which he 1s now Group Leader, A PARADIGM FOR LEARNING by DR. 0. G. SELFRIDGE IuTRouctroN We are proposing here a model of @ process unich we claim can adaptively improve itself to handle certain pattern recognition problems which cannot be adequately specified in advance, Such problers are usual when trying to build a machine to intate any one of a very Large class of hunan data Processing teciniques, A speech typewriter 1s a good exanple of sonething that very many people have been trying wisuccessfully to build for sone tine. We do not suggest that we have proposed a model whitch can learn to typewrite fron merely hearing speech, Pandenaniun does not, however, seo on paper to have the sane kinds of inherent restrictions o inflextoli ity that many previous proposals have had. The baste motif behind our model 1s the motion of parallel processing, This 1s suggested on two grounds: first, 1t ts often easter to handle data ina parallel manner, and, indeed, tt 1¢ usually the nore ‘iatural* manner to handle 1t in; and, secondly, 1t 1s easter to modify an assembly of quasi~Independent modules than 2'nachine all of whose parts interact innediately and in a complex way. We are nov going to apologize for a frequent use of anthroponorphic or DLonorphic terminology. They seen to be useful words to descrive our notions, hat we are describing 1s @ process, or, rather, a model of a process, Me shall not describe all the reasons that led to its particular formulation, Dut we shall glve sone reasons for hoping that 1t does in fact possess 1 flexibility and adaptability that we ascribe to tty THB PROBLEM ENVIRONMENT FOR LEARNING Pandenonium 13 @ model sihich we hope can learn to recognize pattems whick have not been specified, We mean that 1n the following sense: we present to the model exanples of patterns taken fron some set of then, each tine Anfoming the model which pattern we had just presented, Then, after sone ‘ine the model guesses correctly wich pattern has just been presented vetore (84008) st we infom it, For a Lange clags of pattern recognition ensembles there hes never existed any adequate written or statable description of the distinc tions between the patterns. The only requtrenent we can place on our model ts that we want 16 to Denave In the sane way that men observably behave in, In an absolute sense this is a very unsatisfactory definition of any task, but tt may be apparent that Lc is the way in which most tasks are defined for nost men. Lucky Ls he whose Job can be exactly specified In words without any ambigulty or necessary tnferences. The example we shall {lustrate tn sone detatl 1s translating fron manually keyed Horse code into, 8ay, typewritten messages. Now 1t 1s true that code one eats that a dash should be exactly three tines the lengtn of dot and so on, but 1 tums out that tits ts really mostly irrelevant. what matters 1s only what the vast army of people who use Horse code and with whon one 18 going to have to communicate understand and practise when they use It, It turns out that this ts nearly always very different fron school book Morse, In the sane way the only adequate definition of the pattern of a spoken word, or one hand-written, must be In terms of the consensus of the people who are using tt. We use the term pattern recosnition in a broad sense to Include not only that data processing by which Images are assigned to cne or another pattern tn gone set of patterns, but also the processes by which tne patterns and Gata processing are developed by the organiser machine; we generally call this latter "learning". PANDEMONIUM, IDEALIZED AND PRACTICAL We first construct an idealized pandenontun (fig. 1). Each possible pattern of the set, represented by 4 denon in a box, computes nis similarity with the image stmultaneously on view to all of then, and gives fan qutput depending monotteally on that similarity. The decision denon on top makes a chotce of that pattern belcelng to the denon whose output was the largest. ‘thts to an exact correiate of @ commnteacions eyoten wherein given @ recetvet msasye WUD) a's numer of posstote tranontted seseases Wy(2), iat Ho Thess mat, jr deemed tomate been tranomitcety whieh afetatzes Fahey [See (Quen 8 proceaure to opeiman under cereatn conattions), This eeetiatie as iene, the sare of evalowonce ih a sigh pase sce ~ eee to es ehronttes nessge iv anlected shat 1a most ols ar 0 ence me, (94009) DECISION DEMON Figs, Tdealizes Pandenontun, (24009) 518 PICKS THE LARGEST ‘OUTPUT FROM: THE: COGNITIVE DEMONS, WHO INSPECT. THE DATA OR IMAGE, Fach demon may, for exemple, be assiged one letter of the alphabet, so that the task of the Adenon is to shout as loud of the anount of *A-ness* that he sees in the Image.* Now 1t will usually happen that with a reasonable collection of categories ~ Like the letters of the alphabet - the conputa~ tlons performed by each of these ideal cognitive denons will be largely the sane. In many Instances a pattern 1s nearly equivalent to sone logical function of a set of features, each of which 18 individually common to perhaps several patterns and whose absence 1s also common to several other patterns. ‘ie therefor anend our idealized Pandencniun, The anended version - fig. 9 - bas sone profound advantages, chief among wnich 1s 1ts suscept= tility to that kind of adaptive solf-Improvenent that I call learning. ‘The difference between fig. 1 and fig.3 1s that the common parts of the computations that each cognitive denon carries out in fig. 1 nave in fig.3 been assigned Instead to a host of subdenons. At this stage the orgeniza— ton has Tour levels, At the bottom the data demons serve merely to store and pass on the data, At the next level the computational danons or sub- Genons perform certain more or lees complicated computations on the data and pass the results of these up to the next level, the cognitive denons who weigh the evidence, as it were, Each cognitive denon computes a shriek, and from all the shrieks the highest level denon of all, the dectsion demon, merely selects the loudest, ‘THE CONCEPTEGN OF PANDEMONTUM We cannot ab In1tlo know the Ideal construction of our Pandemonium, We try to assure that 1t contains the seeds of self-inprovenent, Of the four levels tn fig.3, all but the third, the suddenons, which compute, are specified by the task, For the third level, therefore, we collect a large number of possible useful functions, elininating a prior! only those which could not concelvably be relevant, and make a reasonable selection of the others, Deine bound by econony and space, We then guess reasonable welghts for then, The behaviour at this point may even be acceptably good, but usually 1t mst be Improved by means of the adaptive changes we are about to discuss, 4 Te 1s posattle also co phage 1 oo that the andenon te conputing the aistance In gone plase of the Image fron sone ideal a; 1t sens to ne unnecessarily platonte to postulate the existence of "ideal" representatives of pattertis, amd, Indeed, there are often cood reasons for not doing so. # see, fof example, Jerone Bruner, "A study of Thinking*, (84008) sis DECISION DEMON COGNITIVE DEMONS ‘COMPUTATIONAL, DEMONS DATA OR IMAGE DEMONS The Evolution of Pandemonium There ere several Kinds of adaptive changes which we will discuss for our ¥ Pandenontun. They are all essentially very similar, but they may be programmed and discussed separately, Te first may be called "Feature Wetghtinan, Although we have not yet specified what the cognitive denons compute, the sole task at present 1s to add a weighted sun of the outputs of all the computational denons or subdenons; the wetghtings will of course differ for (94009) s17 CONTOURS OF SCORE SCORING VECTOR. TESTED, AND THEREFORE SELECTED, Fig.4, First hii climbing technique: pick vectors at ranacn (potnts tn ‘the space), score then, and select the oe that crores nishest. the different cognitive denons, but the welehtings will be the only aifference between thon, Feature welgnting consists of altering the welents assigned to the subdenons by the cognitive denons so as to maximize the score of the entire ‘y Pandenontum, How then can we do this? The Score What we mean by the score here 1s how well the machine 1s doing, we want 1t to do, This presumes that we are monitoring the machine and telling it when it makes an error and so on, and for the rest of the clscusston we shall be assuming that we have avallable sone such minning score, Now at sone polnt we shall be very interested in having the machine run without that Kind of direct supervision, and the question naturally arises wether the machine can meaningfully monitor 1ts ow performanc: Je answer that question tentatively yes, Dut delay discussing 1t till a later section, task (94009) CONTOURS OF SCORE Second hiti-clinbing techalque: pick vectors (iuunber 4) outseores the previous cuss. fen tracing any that’ decrease. the seohe. 1 ope of then © Share rardon Feature Weighting and jill-Clinbing ‘Te output of any cognitive denon 1s so that the compl the subdenens set of Weights for all the cognitive demons over all $ a vector: Now for sone (unknown) set of wetghts A, the betaviour of 1S whole Pandenontun is optimun, and the problen of ature Welghting 1s co fina (94009) 519 Fig,6. General space showing false peaks. one of the false peaks t8'quite isolated fron the main or true peaks thet set, This nay be described as a hill-clinbing protien, We have a space (of A) and a function on the space (the score), which we may consider an altitude, and which ve wish to maximize by a proper search through A. one possible technique 1s to select weighting vectors at random, score then, and finally to select the vector that scored highest (see figs ¥). It wI11 usually, however, tum out to de profitable to takeedvantace of the continuity properties of the space, witch usually exist In some sense, in the following way: select vectors at randon until you find one that scores percept{bly more than the others, Fron this potnt take small random steps in all Gtrections (that 1s, add small random vectors) until you find a direction chat Improves your score. Mien you find such a step, take It and repeat the process, This ts Illustrated in fig. 5, and 1s the case of a blind man trying to clind a hill. Tere may ve, of course, many false peaks on which one may find oneself trapped in such a procedure (fig.6). (24009) 520 The provlen of false peaks In searching teciniques 1s an old and familiar one, In general, one may hope that in spaces of very filgh dimen- Slon@lity toe Interdependence of the components and the score 1s so great as to maxe very unlikely the existence of false peaks completely 1solated fron the main or true peak. It must be realized, however, that this is a purely experimental question that nas to be answered separately for every hill-cLinbing situation, It docs turn out in Nill-clinbing situations that the chotce of starting point 1s often very important, the main peak may be very prominent, but unless 1t has wide-spread foot-hills 1t may Lake @ very Long time before we ever begin to gain abtitud: This may be described as one of the problens of training, nanely, to encourage the machine or organism ta get enough on the foot~nilis so that small changes in his paraneters will produce noticeable improvenent In hts altitude or score. One can describe learning situations where most of the aifficulty of the task 1les in finding any way of Improving one's score, such as learning to ride a unlcycle, where 1t takes longer to stay on for 8 second than it does to laprove that oe second to 2 minute; and others where It 1s easy to do 2 little well and very nard to do very well, such 4s learning to play chess. It 1s also true that often the main peak 13 2 plateau rather than an Isolated spike, That 1s to say, optinal behaviour of the mechanism, once reached, may be rather insensitive to the change of sone of the paraneters. Subdemon Selection ‘The second kind of adaptive change that we wish to incorporate nto our Pandenontun is subdenon selection. At the conception of our denontac assembly we collected sonewhat arbitrarily a large number of subdeions which we guessed would be useful and assigned then weights also arbitrarily. The first adaptive ctange, feature welghting, optinized these welghta, but w have no assurance at all that the particular subaenons we selected are good ones, guddenon selection generates new subdenans for tril and eliminates inefficient ones, that, 1s, ones that do not much help Improve the score. We propose to do this initially by two different. tecniques, witch may be called “mutated fission" and "conjugation". The first point to note 1s that 1t 1s possidls to assim 2 worth to each of the subdemons, It may te done in several ways, and we may, for example, write the worth W, of the 42 denon i so that the worthy denons are those whose outputs are likely to affect most strongly the decisions made. (94009) set Wo assume that feature welghting has already run so long that the behaviour of the machine has been approxinately optimized, and that scores ‘and worthe of machine and 1ts denons have been obtained, Fret we eliminate ‘those subdenons with low worths. Next we generate new subdenons by mutating the survivors and rewelghting the assembly, At present we plan to pick one subdenon and alter sone of his paraneters more or less at Tandon, This will usually require that we reduce the subdenon hinself to sone canonical form so that the random changes we Insert do not have the effect of rewriting the progran so that 1t will not run or eo that 1t will enter a closed loop without any hope of getting out of 1t.+ Besides mutated fselon, we are proposing another method of subdenon improvenent called "conjugation", our purpose here 1s two-fold: first to provide a logical vartety in the functions computed by the subdemans, ‘and, secondly, to provide length and complexity in then. mat we do 1s thist given two Tuseful! subdenons, we generate a new ‘gubdenon whieh 15 the continuaus enalogue of one of the ten nontrivial binary two-variable functions on then, For exanple, the product of two ‘subdenon outputs, corresponding to the logical product, would suggest the Simultaneous presence of two features, The ten non-trivial such functions are Listed in Table 1. A.B Ay B ADB AB ~A>B ~Ay~B “A.B ~Ay~B A.ByASB AMBy~A. B rable 6 W rrivial binary funetions on tno vartables. Jour Pancenontun of an 184 704. The analogues for eobvioie wands of sinuiation (9aco9) 522 Control Adaptation ‘The first two levels of adaptation are directly concerned with innedtate Inprovenent of behaviour and the score, We should also 1lke to Inprove the entire organization, and in the sane way, We shall deal with thls point scnewhat cursorily, belng reluctant to specify things too far in advance of experinent. In prineLplé, we propose that te control operations should thenselves be handled by denons subject to changes Like feature wetehting and subdenon selection, It 1s obviously @ Little more difficult and perhags impossible here to define the usefulness or worth of a particular Qenon, It 1s also clear that 1t will sometimes take much longer to creck the usefulness of some change In some control denon - for example, in one of those which control the mtations in subdenon selection. Furthemorey at this level, some of the denons, presumably, will ve in a position to chanse thenselves, for otherwise We should need another level of possible change, and 0 ons This raises the possibility of irreversible changes, and 1b ts not obvious that all parts of the machine should be subject to adaptive change, But these are Largely heuristic questions, The Evolutionary Process ‘The adaptive changes mentioned above will tend, we hope, to promote a kind of evolution tn our Pandenontun, The scliene sketched 1s really @ natural selection on the processing demons, If they serve s useful function they survive, and perhaps are even the source for otier subdenons who are thenselves judged on their merits. Tt ts perfectly reasonable to conceive of this taking place on a broader scale - and in fact 4t takes place almost Inevitably. Therefore, instead of having but one Pandenoniun We mient have some crowd of then, all fairly alnilarly constructed, and enpLoy natural selection an the crowd of then. Elininate the relatively poor and encourage the rest to generate new machines in their om tnages. Unsupervised Operation 50 far all of the operation of the machine nas been on the basis of ‘constant monitoring by a person who Is telling the machine when tt makes an error. A very valld question 1s whether the machine ean form any independent opinion of 1ts avn on how well 1t 1s doing, 1 suggest that 1t ‘can in the following way: one eriterion of correct decistons will De that they are fairly unaquivocel, that 13, that there is one and only one coguttive denon whose outpat far outshines the rest. Sone running average of the degree to whites this 1s 50 would presumably sonevhat reflect the Score of the machine, Note that 1t would be vital that the machine be trained first to do well enough before 1t 1s left to 1ts om resources and supervision, (94008) 523 ort = bor as \ cue Le La " NSS 4 7 ry 4 Te eavrangcen oaees st sams Red ieeusiee: eae i ae Flg.7, First morse pandenontun, A REAL-LIFE EKAMPLE: MORSE TRANSLATION AS we mentioned before, the entire notion of Pandemonium was concetved as 2 practical way of automatically Improving data-processing for pattem recognition, our initial model task 1s to distinguish dots and dashes in manually keyed Morse code, so tat our Pandemonium can be 11lustrated In fig.7. Note that the functions and behaviour of all demons have been specified except for the computing subdenons. We shall reiterate those spect fications. (1) The decision denon's output 1s "dot! or ‘dasht according as the dot demon's output 15 greater or less than the dash demon' s, (2) me cognitive denons, dot and dash, each conpute a welshted sun of the outputs of some 150 computing subdemons, Initial wetghta we have assigned arbitrarily, but, we Hope, reasonably. (3) The data-nandling demons receive data 1n the form of durations, alternatively of marke and spaces, and they pass then dow the Line, (94009) 524 ourut Os ro 5 Flg.8. Operation of the subdenons da * doy dap ‘The computing subdenons are constructed fron only @ very few operational functions, whieh are carefully non-binary, For exanple the subdenons dotdg and d, dp have their outputs shown in fig.8. The operational functions ToLLow: (1) '=!, tis tinction computes the degree of equality of sone set of vartables (see fig. 8s (2) "4, +4, compute the degree to which sone variable 1s less than or 1s greater than some other variable (see fig. ). (3) ‘max', *max', compute the degree to which some variable ts the largest of an arbitrary set of variables or an arbitrary set of consec~ utive vertables, (4) 10,1, '4j', store the degree to wien the 1th duration has been Adenti fed a8 a dot or dash, (5) ‘4u* computes an average of sone set of variables, (6) Wi" 4s a tantly of tracking means, For exanple, it mlgnt conpute IIc) = Mina) + 2-0 (1) 1Ox!, 'Ax', Oxt 1s the Last duration Identified as 2 dot deg Is the third last duration Identified as a dash, etc. Te above 1s the present functional vocabulary of the ccnputing opere tons for our subdenons, The subdenons thenselves are built with = sinple syntax, For the Initial set, at conception, we merely select a sat of operational funettons and follow then with’ the nur ne particular jata denon, (91009) concnusron what 1 shall present at the meeting In Novenber will be the detalis of the progress of Pandenoniun on the forse code translation problem. The initial problem we nave given the machine 1s to distingileh dots and dashes, ‘hen the behaviour of the machine has improved Itself to the point were Little further improvement seens to be occurring, we shall add three more comitive demons, the symbol space, the letter space, and the word space, Presunably after sone further time this new Pandemonium will settle down to Sone uninprovable state, Then we shall replace the senior or decision-making Genon wth @ row of some forty or so character denons with a new deci sion ‘making denon above then, letting the new cognitive denons for the character denons use a11 the inferior denons, cognitive and otherwise, for their Inputs, It ts prodably also destrable that previous decisions be avatlable for present decisions, so that a couple of new functional operations micnt be added, There need be 1ittle concern about logical etreularity, because we have no requirement for logical consistency, but are merely seeking agreeable Morse translation, How much of the whole program will have been run and tested by Novenber I cannot de sure of, At the present (July) we have had sone fair testing of hill-clinbing procedures, ACKNOWLEDGEMENT I should Like to acknowledge the valuable contributions of atscusstons with many of my friends, including especially M, Minsky, U. Netsser, F Frick and J, Lettvin, (94009) 828 DISCUSSION ON THE PAPER BY DR. 0. G, SELFRIDGE DR J. We BRAY: May I ask Dr, Selfridge what ne thinks of this approach to nis problen? Let x= duration of the last signal, wilen may be a dot or a ash, a signal space, Letter space or wrd space. Let y= 2% if in fact tt was @ dash, 1 1f dot, 0 1f signal space, -1 1f letter space and -2 if word space, 5 ‘To form the polynomial: HAS + AES Age? 6 Age take @ nunber of observations, as he suggests, aid Let the machine learn the code by determining the coefficients Aetc. by simple curvilinear regression. The duration of previous and suvsequent stanals and the Inter pretation given to previous sigals could be added as further variables on the right hand side. MR. Cy STRACHEY: At the end of the paper you promise us sone further information about the latest state of the programe. Could you let us know what this 1s? DA. J. MoCARTHY: I would Like to speak briefly about sone of the advantages of the pandemontum model as an actual model of conscious behaviour. In observing 4 brain, one should make a distinction between that aspect of the behaviour which 1s available consciously, and tnose dehaviours, no doubt equally important, but whlch proceed unconsciously, If one conceives of me brain as a pandemontun - @ collection of denons - perhaps what is going on within the demons can be regarded as the unconscious part of thought, and what the denons are publicly slouting for each other to hear, as the conscious part of thought, DR. DM. MACKAY: DP, Selfridge's ‘pandenontun' has a certain tantly resenblance to a class of mechanism constdered in some earlier papers (ref. 1) (taough I had never suspected its demonic Amplicationst). In one of these, (ref.2) after discussing the general principle that you penalise the unsuccessful, I polnted out that te anount of infomation 1. MACKAY, D. m Mentality dn YachLaes. Proc. Arist, Soc. Supt, 1952) 6. 2, MACKAY, Di M The Bplztenclogical Problen for Autoxata, ecr Jy Mocarchy” ang G. £. Shainah dutonata Studies, Princeton (1996), See also Srit, vy Poychol., dss; 87, 20 aid ddvancenent of Science, 1056, S82, {94008 ser per 'ktox' (to use Dr. geltride's metaphor) 1s very snall unless the probability of success and fatlure are equal. If you have a systen where there are a vast nunber of possibilities to be eliminated by rather feckless trial and error, then of course failure occurs mich more often than success. Jpe solution I suszested was to form a kind of syndicated Leaming process {n'vhter at first latze nuabers of elenents, destined eventually for independence, should 50 48 to reduces Responses A hand, fr example, miuht not at frst have ea cer sepapatel) controllable, but cOuld work clunsily asa wiole. In wat way, you 6 greatly decrease the anount of groping witch 1s necessary before @ successful fay be less, On this principle, gf pursued to the Limit, even the earliest GHiaLe cout nave por-neertgute race OF Seoseeee (nat ooeueas hes fay oa ~~ Given this easy start, staple sel f-orgatsing subroutines can bulld up fairly quickly. as they increase thelr sunber and thelr succes, however, the idea 1s tat the couplings between elenents should gradually be dissolved to Increase te complexity of the problem. If you keep the complexity ineressing step by step with the desree of developsent of successful intemal matening sub-routines, then MLly adaptive behaviour can be enomously nore quickly developed than if the systen starts with the full repertoire to ve explored, My question 1s why Dr. Selfridge nas not Incorporated this prinetple in nis 'pandenontun', so taat each "Kick! could have something nearer to one lt of infomation instead of an almost negligible fraction. DR. P. 0. PRICE (written contribution): Tis 1s a very interesting and stimulating paper. I nave one comment to make, and that 1s on the discus- ston of "Feature Welmting and HIl1-CLinbins", I think that in unis discussion the author has not brought out one important distinction between types of "nlll~climbing* problems ~ that between determinate and stochastic problen, wiereas in a determinate problem the *nill" 1s defined by a single function of many variables S08) @ Fey ereery) tn a stochastic problen 1t 1s defined by the mean value of a mumber of functions: Fix) = f(x) wien Sale) * Fly eeixyi a and @ 18 8 randon vartable wiose value depends on the partiaular trial nade. ow I would have thought that most of the more interesting pattem Fecogii tion problens are essentially stochastic: the objective 1s to make (94009) 528 a machine that w11L obtain as nigh a proportion as poscivle of correct, answers toa series of questions "wat 1s tls pactern?* referring to a randon series of pattems. ‘The autnor has glven an excellent account of sone of te atftteulties of determinate hfll-clinbing and of the techniques for overconing then, DUE stochastic Bili-elimbing presents additional problens, and I think this may vecone very apparent when a pandenoniun 1s vullt to deal wita a practical task In particular, two problens which will need investi sation aret (4) ow Large a sample of trial values f(x) and f(y), Delonglns to two points x and y in the vector space being explored, will be needed to obtain a satistactory estimate of [f(x) - fy] for the purposes of niil-clinbing? (11) how mich Longer will a parttalar smehastic nill-clinving procedure take than the corresponding determinate procedure, and will this ratio prove to be too large, in sone practical Instances, to allow the evolu- tionary developaent of the pandemonium to take place? I siould be very Interested to leam If the author has considered these problens and developed any solutions. The problen of stochastic nill- clinbing 1s of practical Interest outside the context of tis paper, for tnstance In connection with the evolutionary operation of chealesl processes. DR. 0. G SELFRIDGE (1n reply): Since I have been working on Morse code, woten I en doing in addition to working on leaming, I nave aet probably 50 people and when I say that I a working on a machine to do manually keyed Morse code, they say *I will tell you how to do that" and they ten proceea Yo cone up witn sme scheme, Actually, thls 1s the first vine I nave heard tals particular one, whereas I have cone across the otiers many tines, so Dr. Bray should be congratulated on @ new schene for solving Horse code, Let ne assure hin that 1t will not work; we have tried 11, slorse code manually recelved 1s not that simple. The context necessary apparently extends at Least 10 letters on either side - not Just one side, both sides, In fact, the real giestion avout doing this ~ wily T chose Morse code ~ was exactly that 1 had had this Kind of interactions wth each other. You cannot do it by looking at binary functions of the durations, tat 1s exprecsing aurations as binary digits and then looking on binary functions of lots of them. If 1 take 10 Letters on ether side, each has three marks, you should consider the spaces as wall, bat leaving those out, tat 1s three times 20, whlch 1s 60 durations, and if we specify then to one part in $2, that 18 800 Dits, and binary functions of 300 bits cannot be pleked at random. 50 te question 41s to take sone steps in the right direction and then nope te steps are Fight enough so that you wI11 get enough Improvenents, so that tne following steps will get you even more so. Yeatted versions by the author was not available. (84009) I maintain that trere is sone merit in studying sel finprovenent systens and if 1 an going to do that I an going to study systems where 1 mow wat want, rather tan more difficult problens, however attractive they may be to mathenatictans. datnenaticlans Like to work on unsolved problens for te weaver alory, and proviens already solved 1tke the prime nunber theoren are Lert to graduate students. I nave 2 susplelon that John Mecarthy might Later oring up sone inportant points about descriptions, aid here I see my point about solving useful problens, vecaase Horse code 1s a useful problem. ‘The whole question 1s at what level you are woing to deal with descriptions of your dats. In most provlens, especially 1n dealing with the anount of things when people do, binary functions are Just not adequate. For one thing there 18 too much data, and one of the first questions you talk about in the conditioned reflex ~ rengnber that conditional probability assunes that you have already recomisef the stimulus; if you know that the stimilus belongs to one of a small class of functions, it is pretty easy, a8 aman, but mostly tn hunan problens you do not know tis. br. Mackay made sue very Kind renarks. I would go further back than he etd, as he well knows, tn crediting demonology. All of us have sinned in Adan, we Hlave eaten of the tree of the knowledge of good and evil, and the denonological allegory 1s a very old one, indeed. As for iis renarks about the syndicate approach, the evolution of diversative organisation, I think that 1s an extremely accurate and good point. One of the things a species Learns 1s not only how to survive but to have exactly the right variants in tts memers, The reason florseshge crabs have not changed much 1s not Decause they don't adapt; they can adapt, but there 1s no vartation, so they can only adapt very Little, and they are atout as god as they can whereas tere are all kinds of peopl Ta speed aid Infomation theory, I really consider that speed in the classic sense of conputstion 1s so completely irrelevant to this problem - the number of binary operations one does a second ~ that 1t does not interest me very mich though 1t Interests conputer destmers. 1 would rather Like to see lots of operations witen could concelvably ve cone tn parallel, being done seqehtially, because this 1s the only machine we have, and then T hope people will nave machines whteh will work In parallel, so that when I want a machine to do twice as difficult a problen I merely build wiice as Dig 4 machine, instead of letting one machine work vate a8 lone. Dr, Strachey asked about results, and they are roughly as I have Indicated, Inprovenent does take place; the nill-climbing does work. ‘Te sub-denon selection in the progrannes that we lave did not do wiat we hoped, but the sub-denons were effectively throw out, and the ones witch were kept Were largely concentrated around those demons wich relate those ones very close. ‘he most effective demon by Itself, wien worked 99 por cent. of the tine, whic was very surprising, was wnen a mara was the maxima, oF within a just noticable difference of the naxinun of six denons clustered round zero fron minus 3 to plus % I an afrald that that 1s as much as one can say fatrly.

You might also like