.pn 0.ls1.EQdelim $$.EN.ev1.ps-2.vs-2.ev\&.sp 10.ps+4.ceIN SEARCH OF ``AUTONOMY''.ps-4.sp4.ceIan H. Witten.sp2.ce4Department of Computer ScienceThe University of Calgary2500 University Drive NWCalgary, Canada T2N 1N4.sp2.sh "Abstract".ppThis paper examines the concept of autonomy as it pertains to computersystems.Two rather different strands of meaning are identified.The first regards autonomy as self-government or self-motivation.This is developed by reviewing some recent AI research on representing andusing goals, together with physiological, psychological, and philosophicalviewpoints on motivation and goal-seeking behavior.The second concerns the biological independence of organisms which have theability to maintain their own organization in a capricious environment.The advantages of such organisms have been realized recently in a number ofdifferent computer contexts, and the examples of worm programs,self-replicating Trojan horses and viruses are introduced and discussed..bp 1.ls2.sh "Introduction".ppWhat does it mean for a machine to be autonomous?Has any progress been made towards autonomous machines since Grey Walter'sfamous \fIM.\ Speculatrix\fR\u1\d (Walter, 1953)?.[Walter 1953 living brain.].FN1.\ \ for the discerning, or ``tortoise'' for the profane, as its inventortook pains to point out..EFIn a narrow sense it is clear that there has, as evidenced by the evolution ofthe \fIM.\ Labyrinthea\fR species (of which Claude Shannon constructed anearly example) into the fleet-footed trial-and-error goalseeking devices seen in successive generations of the IEEE Micromicecompetition.However, these devices have a predictable course and a predestined end,providing an excellent example of the old argument against artificialintelligence that ``reliable computers do only what they are instructed todo''.In this paper we seek autonomy in some deeper sense..ppIt is not surprising that dictionary definitions of autonomy concentrate onnatural systems.According to the Oxford dictionary, it has two principal strands of meaning:.LB "\fBAutonomy\fR  1.  \fBa\fR  ".NI "\fBAutonomy\fR  1.  \fBa\fR  "\fBAutonomy\fR\ \ 1.\ \ Of a state, institution, etc.NI "\fBa\fR  "\fBa\fR\ \ The right of self-government, of making its own laws andadministering its own affairs.NI "\fBb\fR  "\fBb\fR\ \ Liberty to follow one's will, personal freedom.NI "\fBc\fR  "\fBc\fR\ \ Freedom (of the will): the Kantian doctrine of the Will givingitself its own law, apart from any object willed; opposed to \fIheteronomy\fR.NI "1.  \fBa\fR  "2.\ \ \fIBiol.\fR  autonomous condition.NI "\fBa\fR  "\fBa\fR\ \ The condition of being controlled only by its own laws, and notsubject to any higher one.NI "\fBb\fR  "\fBb\fR\ \ Organic independence.LE "\fBAutonomy\fR  1.  \fBa\fR  "Our interest here lies in practical aspects of autonomy as opposed tophilosophical ones.Consequently we will steer clear of the debate on free will and what it meansfor machines, simply noting in passing that some dismiss the problem out ofhand.For instance, Minsky (1961) quotes with approval McCulloch (1954) that our\fIfreedom of will\fR ``presumably means no more than that we can distinguishbetween what we intend (ie our \fIplan\fR), and some intervention in ouraction''\u2\d..FN2.\ \ This seems to endow free will to a Micromouse which, having mapped themaze, is following its plan the second time round when it finds a newobstacle!.EF.[Minsky 1961 steps toward artificial intelligence.].[McCulloch 1954.]We also refrain from the potentially theological considerations of what ismeant by ``higher'' laws in the second part..ppHow can we interpret what is left of the definition?In terms of modern AI, the first meaning can best be read asself-government through goal-seeking behavior,setting one's own goals, and choosing which way to pursue them.The second meaning, organic independence, has been the subject of major debatein the biological and system-theoretic community around the concepts of``homeostasis'' and, more recently, ``autopoiesis''..ppOur search in this paper will pursue these strands separately.Goals and plans have received much attention in AI, both from the point ofview of understanding (or at least explaining) stories involving human goalsand how they can be achieved or frustrated, and in purely artificial systemswhich learn by discovery.Biologists and psychologists have studied goal-seeking behavior in people,and come to conclusions which seem to indicate remarkable similarities withthe approach taken by current AI systems to setting and pursuing goals.On the other side of the coin, there are strong arguments that thesesimilarities should be viewed with a good deal of suspicion..ppThe second strand of meaning, organic independence, has not been contemplatedexplicitly in mainstream computer science.There have been a number of well-known developments on the periphery ofthe subject which do involve self-replicating organisms.Examples include games such as ``life'' (Berlekamp \fIet al\fR, 1982) and``core wars'' (Dewdney, 1984), as well ascellular (eg Codd, 1968), self-reproducing (eg von Neumann, 1966),and evolutionary (eg Fogel \fIeg al\fR, 1966) automata..[Dewdney 1984.].[Berlekamp Conway Guy 1982.].[Codd 1968 cellular automata.].[von Neumann 1966 self-reproducing automata.].[Fogel Owens Walsh 1966.]However, these seem artificial and contrived examples of autonomy.In contrast, some autonomous systems have recently arisen naturally incomputer software.We examine the system-theoretic idea of ``autopoiesis'' and then look at thesesoftware developments in this context..sh "Goal-seeking \(em artificial and natural".ppIn a discussion of robots and emotions, Sloman and Croucher (1981) note thatmany people deny that machines could ever be said to have their own goals.``Machines hitherto familiar to us either are not goal-directed at all(clocks, etc) or else, like current game-playing computer programs,have a simple hierarchical set of goals, with the highest-level goal put thereby a programmer''..[Sloman Croucher 1981 robots emotions.]They postulate that robots will need \fImotive generators\fR to allow themto develop a sufficiently rich structure of goals; unfortunately they do notsay how such generators might work.To exemplify how goals are used in existing AI programs, we will brieflyreview two lines of current research..rh "Examples of artificial goal-seeking."Those working on conceptual dependency in natural language understanding havelong recognized that stories cannot be understood without knowing about thegoal-seeking nature of the actors involved.Schank & Abelson (1977) present a taxonomy of human goals, noting thatdifferent attempts at classification present a confusing array of partiallyoverlapping constructs and suggesting that some future researcher mightsucceed in bringing order out of the chaos using methods such as clusteranalysis..[Schank Abelson 1977.]They postulate the following seven goal forms:.LB.NPSatisfaction goal \(em a recurring strong biological need.brExamples:  \fIhunger\fR, \fIsex\fR, \fIsleep\fR.NPEnjoyment goal \(em an activity which is optionally pursued for enjoyment orrelaxation.brExamples:  \fItravel\fR, \fIentertainment\fR, \fIexercise\fR(in addition, the activities implied by some satisfaction goals mayalternatively be pursued primarily for enjoyment).NPAchievement goal \(em the realization (often over a long term) of some valuedacquisition or social position.brExamples:  \fIpossessions\fR, \fIgood job\fR, \fIsocial relationships\fR.NPPreservation goal \(em preserving or improving the health, safety, or goodcondition of people, position, or property.brExamples:  \fIhealth\fR, \fIgood eyesight\fR.NPCrisis goal \(em a special class of preservation goal set up to handle seriousand imminent threats..brExamples:  \fIfire\fR, \fIstorm\fR.NPInstrumental goal \(em occurs in the service of any of the above goals torealize a precondition.brExamples:  \fIget babysitter\fR.NPDelta goal \(em similar to instrumental goal except that general planningoperations instead of scripts are involved in its pursuit.brExamples:  \fIknow\fR, \fIgain-proximity\fR, \fIgain-control\fR..LEThe first three involve striving for desired states;the next two, avoidance of undesired states;the last two, intermediate subgoals for any of the other five forms.Programs developed within this framework ``understand'' (ie can answerquestions about) stories involving human actors with these goals(eg Wilensky, 1983; Dyer, 1983)..[Wilensky 1983 Planning and understanding.].[Dyer 1983 in-depth understanding MIT Press.]For example, if John goes to a restaurant it is likely that he is attemptingto fulfill either a satisfaction goal or an entertainment goal (or both).Instrumental or delta goals will be interpreted in the context of theprevailing high-level goal.If John takes a cab to the restaurant it will be understood that he isachieving the delta goal \fIgain-proximity\fR in service of his satisfactionor entertainment goal..ppOur second example of goal usage in contemporary AI is Lenat's ``discovery''program \s-2AM\s+2, and its successor \s-2EURISKO\s+2 (Davis & Lenat, 1982;Lenat \fIet al\fR, 1982)..[Davis Lenat 1982.].[Lenat Sutherland Gibbons 1982.]These pursue interesting lines of research in the domains ofelementary mathematics and VLSI design heuristics, respectively.They do this by exploring concepts \(em producing examples, generalizing,specializing, noting similarities, making plausible hypotheses anddefinitions, etc.The programs evaluate these discoveries for utility and ``interestingness,''and add them to the vocabulary of concepts.They essentially perform exploration in an enormous search space, governedby heuristics which evaluate the results and suggest fruitful avenues forfuture work..ppEach concept in these systems is represented by a frame-like data structurewith dozens of different facets or slots.The types of facets in \s-2AM\s+2 include.LB.NPexamples.NPdefinitions.NPgeneralizations.NPdomain/range.NP