Viewing the History of Science as Compiled Hindsight Lindley Darcien

hen I was invited to give a should be aware of the need to pre- talk on AI and the history of serve records of their activities. This article is a written version of an science, I declined the sugges- Pamela McCorduck’s entertaining and invited talk on (AI) tion that I discuss the place of AI in journalistic account of AI, Machines and the history of science that was pre- the history of science. The field is too Who Think (1979,) will be the first of sented at the Fifth National Conference on new, developing too rapidly, to try to many chronicles of this field. Profes- Artificial Intelligence (AAAI-86) in assess its place in the history of sci- sional historians will soon follow the Philadelphia on 13 August 1986 Included ence just yet. Some aspects of AI are journalists. is an expanded section on the concept of an abstraction in AI; this section responds applied and more like engineering The difficulty in preserving pro- to issues that were raised in the discussion than science. Some aspects, those in grams and the hardware on which which followed the oral presentation which AI researchers aim at modeling they run will make the historians task The main point here is that the history human intelligence, merge into cogni- even more difficult than the usual of science can be used as a source for con- tive science. Those researchers trying problem, for example, of coping with structing abstract theory types to aid in to understand intelligence, human or scientists who cleaned out their files solving recurring problem types. Two theo- otherwise, are forging a new area that and threw away their notebooks. ry types that aid in forming hypotheses to might or might not ultimately be Those of you working in the field solve adaptation problems are discussed: called scientific. I don’t believe it is should be careful about preserving selection theories and instructive theories. profitable to argue over whether a material. Consider donating your Providing cases from which to construct theory types is one way in which to view field deserves the honorific title of papers to university or corporate the history of science as “compiled hind- “science.” Each field needs to develop archives, and be aware that future gen- sight” and might prove useful to those in its own methodologies, which are erations will likely want to trace your AI concerned with scientific knowledge shaped by the needs and problems steps of thinking and implementing. and reasoning within the field, without succumbing One scientist whose work I have tried to the disease of physics envy. If some to trace cleaned out his files every five of the many characteristics attributed years. Another researcher destroyed to science are useful within a new his notebooks when he retired. Be- field, then they should be considered. cause we cannot date a key discovery Some fields, for example, might find on the basis of information in this re- controlled experimentation useful; searcher’s notebooks, which are the others might strive for quantitative best historical source, we are left with results; and still others might find the- trying to reconstruct what he said in ory-driven methods of use. However, his classes during this period to see if adopting methods simply to obtain he discussed as yet unpublished the label scientific and not because results in his lectures. Would you they are demanded by the problems want to count on your students’ lec- seems to me to be misguided. ture notes to substantiate your claim Although I believe it is too early to that you made a discovery? Certain try to assess AI’s place in the history large projects, such as the space tele- of science, it is not too early to begin scope, hire a historian at the outset so preserving records for the historians that appropriate records can be kept who will be writing the history of AI. and interviews done as the project AI is an exciting and rapidly develop- proceeds. Scientists interviewed many ing field. Those persons shaping it years later might or might not accu-

SUMMER 1987 33 some remarks about AI that are worth recalling. Schank discussed the some- Although I believe it is too early to try to assess times tenuous connection between AI AI’s place in the history of science, it is not too and the rest of computer science. He said that AI is a methodology applica- early to begin preserving records for the historians ble to many fields other than computer who will be writing the history of AI. science because it is “potentially, the algorithmic study of processes in every field of inquiry. As such the future rately remember what happened and (not the social or artificial]. I am inter- should produce AI anthropologists, AI when. Having a historian involved in ested in scientific knowledge and rea- doctors, AI political scientists, and so the early stages is a good way to record soning. My topic deals with the ways on” (Schank 1983, p. 7). I would add to AI work on large projects. in which AI and the history and phi- this list AI historians and philosophers losophy of science can profit from The Babbage Institute and the of science. Smithsonian Institution have strong interacting. Successes already exist Students educated in computer sci- with such an interface, and I would interests in the history of computing, ence as undergraduates who then spe- like to suggest additional points of but they have done little yet on the cialize in AI as graduate students learn history of AI. Arthur Norberg of the potential contact. a lot of skills, maybe some mathemat- Babbage Institute told me that they When I was invited to speak, I was ics, but they might be exposed to very will be happy to advise anyone with asked to be entertaining and provoca- little actual knowledge, scientific or tive. Because I am normally rather archival material in AI about appropri- otherwise, or to problems examined by ate places for storage. A major exhibit matter-of-fact and, I guess, sad to say, a researchers in other fields. I think this on the information revolution is bit dull, such a task isn’t especially lack of training in knowledge-rich easy for me. AI has some very bright, scheduled to open at the Smithsonian fields accounts in part for so many of Institution in 1990. Currently, nothing clever, and entertaining people who the toy problems and games that have on AI is scheduled to appear in this give very lively talks. I can’t hope to been the domain of much AI work. Al match their liveliness and wit. Also, exhibit. efforts that address real, rather than In contrast, the Computer Museum being provocative requires one to take toy, problems, especially scientific in Boston has an exhibit entitled, dogmatic stands so that others will ones, are of much greater interest to “Artificial Intelligence and Robots,” disagree. Philosophical training has those researchers in other fields sharpened my ability to see both sides opening in the summer of 1987. Oliver attempting to make connections with of an issue and dulled my ability to be Strimpel is the curator working on the AI. exhibit. The museum already has dogmatic. However, in an effort to sat- The work in expert systems has Shakey and other artifacts and is inter- isfy this request, let me attempt to been all for the good in getting Al ested in acquiring still more. It is also make a few provocative remarks before researchers involved with real-world interested in collecting programs, getting on to the serious--and I hope knowledge and reasoning. Some Al either in listings or in other usable not too dull--points. systems have been driven by researcl forms, for their archives. For the To give an overview of what is com- problems and needs in other areas exhibit, they need rugged, state-of-the- ing, first I will make some provocative, including science. Such success art AI systems which might be of introductory remarks. Then, I will dis- requires those involved in AI to learr interest to the public and which could cuss fruitful interactions between AI about other fields and become con withstand museum use. and the history and of sci- cemed with the research problems ant The AI field is fortunate to have its ence. Next will come a discussion of reasoning strategies in various conten key founding fathers still alive. It is an important idea developed in AI, fields. Graduate programs in AI shoulc certainly time for professional histori- namely, abstraction. This idea will be have less training in computer science ans--those trained in oral history applied in a discussion of abstractions More of the students’ time should bc methodology--to begin collecting that I have found by analyzing the his- spent in research fields where AI tech audiotapes and videotapes, deciding on tory of science as well as an illustra- niques can be applied and where the key historical questions that need to tion involving two types of theories needs of these fields can drive the be asked, and filling in the gaps in doc- from the biological sciences--selection development of new areas of AI. The umentation. The American Associa- and instructive theories. The conclu- history of science has been marked b! tion for Artificial Intelligence might sion is a summary of the compiled successful interfield interactions lead wish to fund such a project, just as hindsight provided by the examples ing to new developments (Darden ant many of the scientific societies have discussed and suggests ways in which Maul1 1977). The potential exists fo done. AI and the history of science can con- the AI field to be a part of many sue Although the history of AI is an tinue to interact. cessful interfield bridges, and a interesting topic, it is not the focus of Schank suggested, what would the] my discussion today. It is not my area AI and Other Fields mark a successful contribution in A of research. I am a historian and a In a provocative article in AI would be developments that are appli philosopher of the natural sciences Magazine in 1983, Roger Schank made cable to a wide range of areas (Schanl

34 AIMAGAZINE 1983, p.7). science. Thus, it is outside the sphere Another case of AI researchers lack- of the . Some ing content knowledge but attempting work has been done using the social to apply their powerful skills is the history of science in AI and would, current work on commonsense rea- thus, be located in the intersection of soning. As someone working on the the history of science with AI outside relations between the history and phi- the philosophy of science, as shown in losophy of science and AI, I don’t find figure 1. A possible example is Korn- the work on common sense of particu- feld and Hewitt’s (1981) effort to build lar interest. Common sense is a very a parallel system that is analogous to a vague term. The eclectic list of topics scientific community of researchers pursued in AI under this rubric that are working independently on the demonstrates a lack of clear focus. same scientific problem. Furthermore, commonsense notions about the natural world are often Figure 1: Interfield Relations wrong. Science has spent hundreds of he philosophy of science can (and years cleaning up commonsense much of it has) involve logical GEN) (Friedland and Kedes 1985) aided misconceptions and improving loose analyses of science without atten- in the planning of experiments. One of methods of reasoning. tion to the historical development of the most interesting pieces of work in Much of Aristotle’s world view was scientific ideas. The work of Glymour, AI from a science historian’s point of a commonsense one: objects stop mov- Kelly, and Scheines (1983) on imple- view is the work by Langley, Brad- ing if no mover is pushing them; the menting the testing of scientific shaw, and Simon (1983) and others on earth is at the center of the universe hypotheses in a logical form is an Baconian induction. They developed and does not move, and the stars example of philosophers of science methods for finding patterns in numer- revolve around the earth; animal (those who aren’t concerned with the ic data. Another of their systems, species perpetuate themselves etemal- history of science) making use of AI GLAUBER, can discover certain kinds ly and act according to teleological techniques. Their work would be of qualitative laws (Langley et al. goals; substances have essential and located in the area of figure 1 in which 1983). These methods are sufficient for accidental properties. Science, or natu- AI and the philosophy of science inter- rediscovering numerous empirical ral philosophy as it was called until sect outside the history of science. laws, such as the laws of chemical the nineteenth century, has served to The history of science and the phi- combination discovered in the nine- correct these and many other com- losophy of science interact in the teenth century. Implementing induc- monsense misconceptions. Thus, sci- interdisciplinary area of the history tive-reasoning strategies in order to ence is an important storehouse of and philosophy of science (HPS). The rediscover past scientific results has knowledge about the natural world. aim of HPS is to understand reasoning provided fruitful insights into scientif- Storing this knowledge and finding in the growth of scientific knowledge. ic reasoning methods. But data-driven, ways of making it readily accessible to HPS researchers make use of extensive bottom-up, inductive methods have AI systems is an important task. As a case studies from the history of sci- their limits, as Langley’s group real- philosopher of science, I find efforts to ence in order to understand scientific ized (Langley et al. 1986). encode and use scientific knowledge of change. I argue that AI and HPS have Consider a simplistic division of sci- much more interest than the attempts and can continue to interact to their entific knowledge into data, empirical to make the vague concept of com- mutual benefit; important work has generalization, and explanatory theory. monsense knowledge precise. (For fur- been and can continue to be done at Inductive methods can be seen as a ther discussion of the view that episte- the intersection of AI and HPS, the way of getting from data to empirical mology should be the study of scientif- center area in figure 1. jj generalizations but not to explanatory ic, rather than commonsense, knowl- theories. Explanatory theories require edge, see Popper 1965, p. 18-23.) concepts not in the data themselves in Having now provocatively suggested AI and the order to explain these data (philoso- that AI researchers could profit from phers of science have argued, for exam- working on scientific problems rather History of Science ple, Hesse 1966). A method for produc- than toy ones and dogmatically assert- Thus far, I have discussed science as a ing explanatory hypotheses is the ed that science has cleaned up com- body of knowledge. However, science hypothetico-deductive method: a monsense misconceptions, I will try can also be characterized by its hypothesis is guessed, and a prediction to turn to claims of greater substance. methodologies, which themselves is then derived from it. The philoso- The history of science and the philoso- have improved over time. Some meth- pher of science Karl Popper advocated phy of science can be related in vari- ods used in science have already this method and stressed the impor- ous ways (see Figure 1). Much of the proved fruitful in AI. Experimentation tance of falsifying instances to elimi- history of science now being written is is often seen as a key method in sci- nate erroneous guesses (Popper 1965). not focused on scientific ideas but on ence. The early work on using AI tech- The hypothetico-deductive method the social and institutional aspects of niques in molecular genetics (MOL- corresponds--or perhaps even gave rise-

SUMMER 1987 35 to the methodology in AI called “gen- generalizations in another instance. in AI (Genesereth 1980; Greiner 1985). erate and test.” The key component Hanson’s schema for retroduction An example of a common abstraction becomes the hypothesis generator. was the following: (1) surprising phe- is a tree structure; instantiations The idea from the philosophy of sci- nomena, pl, p2, p3, . . ., are encoun- include the Linnaean hierarchy in bio- ence that generation of hypotheses is tered; (2) pl, p2, p3, . . ., would follow logical taxonomy and the organization- by unconstrained guessing is not very from a hypothesis of H’s type; (3) al chart of a corporation. helpful in trying to implement a therefore, there is good reason for elab- The concept of an abstraction has hypothesis generator. The DENDRAL orating a hypothesis of the type of H not been well analyzed in AI. Creative work [Lindsay et al. 1980) was an (Hanson 1961, p. 630). new concepts usually start out fuzzy implementation of a generate-and-test Hanson had little to say about what and ill-defined and are used in different methodology for forming hypotheses constituted a type of hypothesis and ways. Such early vagueness is useful within a theoretical context. The gave only a few examples. For for allowing a new concept to “float” to hypotheses could be exhaustively instance, he labeled Newton’s law of an appropriate point, for allowing an specified. However, the hypothesis gravitation an inverse-square type of exploration of its potential areas of space was so large that a key problem application. The use of abstraction is was to find constraints to add to the still developing in AI, and this analysis generator to restrict the search of the TWO powerful ideas is not meant to be definitive. However, hypothesis space. Use of the generate- some attempt to analyze the concept and-test method provides a real chal- developed by AI of abstraction and to distinguish it lenge when the hypothesis space is not researchers in a from other concepts, such as general- well-defined--guessing new theories is ization and simplification, is useful not something that computers are computationally useful here. good at. way are abstraction and Abstraction can be considered both Philosophers of science have criti- a process (which will here be called cized the lack of insight provided by instantiation. “abstracting”) and a product of this pro- the hypothetico-deductive method cess (which will be called an “abstrac- that hypotheses are to be guessed. N. theory. Despite its lack of develop- tion”). Abstraction formation involves R. Hanson (1961), drawing on the ment, Hanson’s idea has appeal: find loss of content. Consider a source and work of Charles Pierce, argued for an types of hypotheses proposed in the the process of abstracting from it. The alternative to the hypothetico-deduc- past, and analyze the nature of the abstraction thus produced is, in a tive method called “retroduction” or puzzling phenomena to which the type sense, simpler than its source. Hence, “abduction.” Abduction is a method of applied. Use this “compiled hindsight” abstracting is one way of simplifying. plausible hypothesis formation. It is in future instances of theory construc- Furthermore, because it has less con- beginning to receive attention in AI. In tion. I think AI can help in the devel- tent than the source, an abstraction their recent AI textbook, Charniak opment of Hanson’s vague suggestions, can apply to a larger class of objects, of and McDermott discussed abduction for instance, by applying the concept which its source is one member. In in the logical form usually referred to of an abstraction to devising computa- this sense, abstracting is like generaliz- as the “fallacy of affirming the conse- tionally useful theory types. ing. quent”: The simplifying and generalizing If a, then b aspects of abstracting are reflected in b The Concept methods for forming abstractions. Therefore, a Although it is possible to form an of an Abstraction abstraction from a single instance, it is Of course, b could result from some- One of the marks of fruitful interfield easier to consider the case of taking thing other than 0. However, if one interactions in science is that each two similar instances and forming the already has the reliable generalization field contributes to developments in common abstraction of which they are that a causes b, one can plausibly con- the others. Thus far, I have been indi- both instantiations. The most straight- jecture that if b occurred, then a cating how ideas from the history and forward case is to take two instances, would explain its occurrence (Chami- philosophy of science have been and delete their differences, and consider ak and McDermott 1985, ch. 8). can be useful in AI. I would also like to their commonalities as the abstrac- Sophisticated versions of abductive suggest how concepts and methods tion. With F, G, and H interpreted as inference are being developed by Reg- developed in AI might be useful in properties of an entity a, (F(a) & G(a)) gia, Nau, Wang, and Peng (1985), as fully developing methods for hypothe- and (F(a) & H(a)) give the abstraction well as Thagard and Holyoak (1985). sis formation in science. F(a). For example, “the deltoid muscle Hanson’s schema for retroduction Two powerful ideas developed by AI has actin, and the deltoid muscle has was more complex than the simple researchers in a computationally use- myosin” and “the deltoid muscle has logical form discussed by Charniak ful way are abstraction and instantia- actin, and the deltoid muscle functions and McDermott. Hanson was con- tion. The analysis of analogy as two as an abductor” abstract to “the deltoid cerned with the discovery of new gen- structures that share a common muscle has actin”. In this case, content eral theories, not merely invoking past abstraction is computationally useful is lost in abstracting: G(a) and H(a)

36 AI MAGAZINE cannot be retrieved from the abstrac- “Whiskers has teeth and a white coat,” tion. Thus, the abstraction is simpler and “Sandy has teeth and a calico coat” than its instantiations. are generalized to “all cats have teeth.” I don’t know if this method of Generalization (in the sense of univer- abstracting was developed historically sal quantification) occurred in con- from the famous “axiom of abstrac- cluding “all cats have teeth,” and tion” in set theory: “given any proper- abstraction occurred in the loss of ty, there exists a set whose members information about coats (see HarrC are just those entities having that 1970, for another discussion by a property” (Suppes 1972, p. 6). No spe- philosopher of science of abstraction cific mention is made of the loss of formation). content here, but the focus on one Another method for abstracting pro- property to the exclusion of others can duces abstract formulas that are gener- involve abstracting from numerous al in the “typically ambiguous” sense. properties to one. This formulation of Constants are replaced by variables: abstraction in set theory shows why F(a) has the abstraction F(x). The abstraction and generalization are changes from (F(a) &. G(a)) to F(a) to often used synonymously. Polya’s F(x) involve progressing to higher lev- work on problem solving has been els of abstraction. For example, “the Stephen Nelson-Fay Photograph influential in AI. In his set of defini- deltoid muscle has actin, and the del- tions, he does not list abstraction but toid muscle has myosin” abstracts to A robot family portrait taken at the Computer Museum in Boston. defines generalization in the following “the deltoid muscle has actin”, which way: (‘. . , passing from the considera- abstracts to “x has actin”. It is easy to schema becomes “overcome a target by tion of one object to the consideration see that AI systems would have no dif- of a set containing that object; or pass- ficulty in forming abstractions by a force.” “Overcome” is a creative ing from the consideration of a dropping conditions or changing con- abstraction for “destroy” and “capture.” A system that could form such cre- restricted set to that of a more com- stants to variables given appropriate ative abstractions would have to have prehensive set containing the restrict- rules for doing so. ed one” (Polya 1957, p. 108). In difficult cases, forming an more semantic content than one which operated merely by replacing Quine distinguishes two kinds of abstraction can involve more than constants with variables. The concept generality that help explicate the con- merely dropping parts or replacing of a schematic structure from which cept of abstraction. The first kind of constants with variables. New, specific content has been omitted is generality is what he calls the “typi- abstract semantic concepts might have often used synonymously with cally ambiguous,” namely, the to be introduced. Forming abstractions abstraction in AI (see Zeigler and Rada “schematic generality of standing for in such cases might involve creativity 1984 for a discussion of this sense of any one of a lot of formulas.” F(a) is in finding the appropriate concepts and abstraction]. typically ambiguous because numer- terminology. This method might be The idea of multiple levels of ous substitutions can be made for F difficult to implement. Suppose abstraction is also important in AI. and for a. This meaning of generality instead of abstracting from “the deltoid Scientists sometimes make use of is close to the usage of abstraction in muscle has actin” to “x has actin,” one mathematical models at one level of AI. Second, he discusses the generality wished to abstract to “abductor mus- of universal quantification, which cles have actin.” The AI system would abstraction, but multiple levels of involves “quantifying undividedly over then have to have an “is a” hierarchy abstraction, with each increasingly abstract, is not a common idea in sci- an exhaustive universe of discourse” representing the knowledge that the ence or the history and philosophy of (Quine 1969, p. 266). For example, deltoid muscle is an abductor muscle. moving from “some muscles contain To move to the next abstract level, the science. In AI it is easy to consider actin” to “all muscles contain actin” is system would substitute the class con- dropping detail as one proceeds upward in an abstraction hierarchy. Forming a step of universal generalization. It cept for the individual. Some guidance the higher-level, abstract concepts would probably not be called abstrac- would be needed to know if the proper- while building a knowledge base can tion because no content is lost (with ty of the individual (in this case “has be a creative process for the knowledge the possible exception of the loss of actin”) applies to the class at the next engineer. The AI researcher might the existential claim that some mus- level in the hierarchy (in this case, it have to add new abstract concepts in cles exist). This example is an instance does--abductor muscles do have actin). order to put the knowledge into a of inductive generalization, not The formation of schemas has been knowledge representation system with abstraction. discussed by cognitive psychologists. multiple levels of abstraction. For In forming a general class, however, Gick and Holyoak (1983) discuss the example, the concept of a “molecular one can omit individual differences to example of forming a schema to repre- switch” was developed at a high level get the class concept, which involves sent two analogous situations, such of abstraction in the MOLGEN knowl- an abstracting step. For example, con- as”destroy a tumor by radiation” and edge base, and it has proved of interest sider forming the class concept of cats: “capture a fortress by an army.” The

SUMMER 1987 37 to molecular biologists who hadn’t For example, in devising a plan to “buy gene into a bacterial plasmid, a splic- considered their material at such an a piano,” the highest level might be ing plan, with its steps of choosing abstract level (Friedland and Kedes “locate piano” or “get money.” Lower- appropriate restriction enzymes and SO 1985; Karp 1985, personal communica- level details such as “drive to store” on, could be retrieved and instantiated tion). Progressive refinement by suc- would be omitted until the later for the particular case. This method is cessive instantiations of multiple lev- stages. If the higher-level, critical goals similar to Schank and Abelson’s (1977) els of abstraction is now a common cannot be satisfied, then the lower- idea of stored scripts to provide expec- idea in AI, but it is not commonplace level details need never be considered tations in story understanding. Skele- in other fields. I will show how it can by the planner. tal plans and scripts are stored abstrac- be a method for producing alternative Korf summarizes: tions based on compiled hindsight. hypotheses. Thus, to summarize, abstraction for- The value of abstraction is well- mation involves loss of content. The known in artificial intelligence. loss makes the abstraction simpler The basic idea is that in order to than its instantiation(s). Also, the Forming the higher- efficiently solve a complex prob- abstracting process might produce an level, abstract concepts lem, a problem solver should at abstraction that has a larger class of first ignore low level details and instantiations than the original one(s) while building a concentrate on the essential fea- from which it was formed; in this knowledge base can be a ture of the problem, filling in the sense, abstraction is like generaliza- details later. The idea readily gen- tion. Multiple levels of abstraction can creative process for the eralizes to multiple hierarchical be formed by continuing to apply knowledge levels of abstraction, each focused abstracting methods at each level to on a different level of detail. Em- produce the next higher one. Several engineer. pirically, the technique has prov- methods for forming abstractions have en to be very effective in reducing been discussed, including ( 1) dropping the complexity of large problems content; (2) replacing constants with Abstraction has been used in AI sys- (Korf 1985, p. 7). tems to do problem solving and plan- variables; and (3) forming new, abstract ning. Korf (1985) claims that the first Three methods for forming abstrac- semantic concepts. A fourth method of explicit use of abstraction in AI was in tions from instances have been dis- abstraction, which is most important the planning version of the general cussed: (1) dropping conditions, (2) in considering applications from the problem solver (GPS) developed by replacing constants with variables, and history of science, is the following: a Newell and Simon in 1972. A discus- (3) finding semantically rich concepts schema that has resulted from drop- at a higher level of abstraction that sion of GPS in the Handbook of Artifi- ping detail and preserving underlying cial Intelligence makes this use clear: capture the meaning of the lower-level structural or functional relations concepts. What guides the dropping of among component parts. GPS planned in an abstraction detail depends on the purpose for space defined by replacing all log- which the abstraction is being formed, ical connectives by a single on the particular problem-solving con- abstract symbol. The original text. A given entity or even two simi- History of Science as problem space defined four logi- lar entities might be abstracted in dif- a Source of Abstractions cal connectives, but many prob- ferent ways depending on which fea- Except for the development of math- lem-solving operators were appli- tures are the focus of current investi- ematical models, which can be viewed cable to any connective. Thus, it gation (see Utgoff 1986 for a related as going up one level of abstraction, could be treated as a detail and discussion of bias in learning and little effort has been made to extract abstraction out of the formula- Kedar-Cabelli 1985 for a discussion of useful abstractions from the history of tion of the problem (Cohen and the role of purpose in forming analo- science that can be instantiated in new Feigenbaum 1982, p, 518). gies). problem situations. The thesis to be Another early example of the use of Thus far, the discussion has focused argued here (philosophers like to argue abstraction in AI was the work of on forming abstractions from for theses) is this: the history of sci- Sacerdoti (1974) in ABSTRIPS, a plan- instances. However, some AI systems ence can serve as a source of compiled ning program. do the opposite; namely, they retrieve hindsight for constructing useful ABSTRIPS plans in a hierarchy of available abstractions and instantiate abstractions. These abstractions will abstraction spaces, the highest of them in a given situation. Friedland’s be of theory types or other recurrent which contains a plan devoid of all (1979) version of MOLGEN stored patterns and processes in the natural unimportant details and the low- skeletal plans containing steps at a world. They might prove useful in new est of which contains a complete high level of abstraction for experi- problem situations--when these situa- and detailed sequence of problem- ment planning in molecular genetics. tions are of a recurring problem type-- solving operators (Cohen and In a given problem, the appropriate for which an available abstraction Feigenbaum 1982, p. 517). skeletal plan could be retrieved and exists. A presupposition behind this instantiated. For example, to splice a thesis is that recurring patterns and

38 AI MAGAZINE processes occur in nature. Recurrent look for abstract characterizations of tion theories that don’t have a step for types of theories that have proved use- biological theories and the reasoning reproduction. Once Darwin formed his ful in the history of science are worth that could produce them. (Although I theory of natural selection, selection analyzing for common features and have begun trying to implement rea- as a type of theory became available. It putting into abstract form. Such ab- soning sufficient for rediscovering a was a product of hundreds of years of stractions can then serve as compila- theory in genetics, this implementa- biological thought and represented a tions of the hindsight gained from tion is very much in the formative departure from commonsense observa- their study. Preserving such hindsight stages. Its discussion is appropriately tions so radical that some still resist becomes especially important when left for technical sessions in future this theory. theories are counterintuitive, and sci- years (see Darden and Rada forthcom- Once Darwin formulated a selection entists have produced insights that ing, for a preliminary sketch). I will theory to explain species change, the serve to correct naive, commonsense now discuss an example of a problem theory type was available for other views. If this pattern is encountered type and a recurrent theory type that instances of theory construction. An again, it will be computationally effi- has resulted from searching for useful example comes from immunology. cient to have an abstraction available abstractions which can perhaps be Antibodies that are adapted to elimi- rather than to have to rediscover it. implemented. nating invaders in the body need to be Discussions at the Analogica ‘85 An adaptation problem involves generated; thus, formulating a theory conference held at Rutgers University, explaining how something comes to fit of antibody production involves solv- New Brunswick, New Jersey, showed something else, or how two things ing an adaptation problem. In improv- that an important next step in the cur- change over time so that one comes to ing a theory for antibody formation in rent work on analogy is to compile be adapted to the other. Two different the 195Os, Burnet specifically used an instances of useful abstractions. Some theory types have been proposed his- analogy to natural selection. He pro- of the individuals working on analogy torically to solve adaptation problems: posed the clonal selection theory: cells have specifically focused on scientific selection theories and instructive the- vary according to the antibodies they examples (Darden 1980; Darden 1983). ories. produce, and a large amount of diversi- Gentner (1983) discussed central-force An abstraction of selection theories ty is present in the naturally circulat- systems, such as the solar system and has the following components: varia- ing antibodies. Selective interaction the Bohr model of the atom. Greiner tion, interaction of the variants in a with invaders occurs. Those cells (1985) constructed an abstraction for constrained environment, and the which produce the appropriate anti- flow problems, such as water and elec- resultant perpetuation of the fit vari- bodies reproduce in clones and produce tric flow. Thagard and Holyoak (1985) ants (reproduction or amplification). an amplified level of the particular investigated wave-type theories, such This theory type has only been antibody until the invader is eliminat- as water waves and sound waves. The abstractions being developed are com- putationally useful in doing analogical reasoning and problem solving when The history of science can serve as a source of appropriate problem types arise. The compiled hindsight for later stages of the work on Baconian constructing useful abstractions. induction were not specifically fo- cused on abstractions in analogical reasoning; nonetheless, they involved available since the middle of the nine- ed (if all goes well). Both Darwinian reasoning in discovering types of theo- teenth century when Darwin proposed natural selection and clonal selection . ries: compositional theories, and par- his theory of natural selection to have fared well in subsequent tests and ticulate theories (Langley et al. 1986; explain the origin of new, adapted with further developments in their Zytkow and Simon 1986). species. Darwin amassed evidence to fields. Bringing a computational AI per- show that variation occurs in nature. Selection as a type of theory can spective to case studies in the history He argued that the environment can- now be seen as a historically success- and philosophy of science provides a not sustain all the organisms which ful means of biological theory con- new focus. The task is to find abstract are produced, so there is a struggle for struction for adaptation problems. It patterns and processes that can be existence. The adapted variants tend to can now be used as an abstract pattern implemented and used in more cases. survive. In natural selection, survival for forming new theories to solve new into the next generation by reproduc- adaptation problems. Abstractions for tion is what ultimately counts for a An interesting attempt to formulate variant’s success. Reproduction can be a new selection theory is Edelman’s Selection and Instructive included at a lower level of abstraction “group-selective theory of higher brain Type Theories as one of several ways to instantiate function” (Edelman and Mountcastle Bringing this computational perspec- the perpetuation of the fit variants. By 1978). As far as I know, it has not yet tive to my own area of the history and removing it from the higher-level been confirmed or disconfirmed. The has led me to abstraction for selection theories, the unit of variation is a group of brain abstraction applies to additional selec- cells. Groups of cells interact with sig-

SUMMER 1987 39 nals such that some are amplified and (Edelman and Mountcastle 1978, p. develop criteria for choosing which others are not. This amplification can 56). abstraction to instantiate; instantiate occur by either positive or negative se- However, in a known future--name- the appropriate abstraction in the cur- lection of the groups of cells. Because ly, a stable environment--developing a rent adaptation problem situation to the theory is still in its formative generator to make just adapted forms provide new hypotheses; test the stage, alternative ways of instantiating might well be the better strategy. One hypotheses; and refine the criteria for the abstraction provide competing might, for instance, apply these types applying the abstraction in light of hypotheses to be tested, such as this of theories to the adaptation problem successes and failures. case of positive or negative selection. of making a part in a factory that is A general research program emerges As other theories are formed that adapted to another part. An instructive from the thesis that the history of sci- instantiate the abstraction, then alter- process would be better than a selec- ence provides compiled hindsight: natives at a slightly lower level of tive one if the kind of part to be pro- study the history of science to find abstraction can be elaborated to pro- duced will be stable over a sufficiently recurring problem types and theory vide alternative hypotheses. It will be long period of time, and a means exists types, devise computationally useful interesting to see the fate of this new for communicating to the mechanism abstractions for them, and build AI selection theory. generating the part the specifications systems to use such compiled hind- Another type of theory predated the part should meet. In another sight in new problem situations. selection theories as an alternative for example, debates in evolutionary epis- Important tasks in the coming years solving adaptation problems-instruc- temology have raised the question of will be to put scientific knowledge tive theories. Instead of a random set whether humans blindly generate into a form that can be computational- of variants, information is used to gen- alternatives when forming new con- ly useful and to devise reasoning erate variants that are already fit. Neo- cepts and then select them or whether strategies which can be implemented Lamarckian inheritance of acquired a directed, instructive process better to produce and make use of scientific characteristics is an example. A new captures human knowledge acquisi- knowledge. Successful use of the com- form is generated in a changed envi- tion (see Campbell 1974 and Thagard piled hindsight from the history of sci- ronment to be adapted to the con- forthcoming for discussions of evolu- ence could contribute to this task and straints in the environment. For exam- tionary epistemology). show the usefulness of interfield inter- ple, the giraffe sees the leaves high in In considering the hindsight provid- actions between the history of science the tree and grows a longer neck to ed by these historical cases, considera- and AI. reach them, which is then inherited by tion of the incorrect theories, as well Earlier at the conference, I had the baby giraffe. No wasted variants as the correct ones, has proved useful. lunch with Doug Lenat, from whom I are produced, and no selection is nec- Scientists often consider disproved have taken the phrase “compiled hind- essary to produce the adaptation. theories in a scientific graveyard, not sight” (Lenat 1983). I complimented Although this theory eliminates the worth a second glance by those Doug on his ability to give very enter- waste of producing unadapted vari- researchers pushing ahead at the fore- taining talks. He said one rule is to ants, it necessitates a powerful genera- front. However, in these cases, incor- close with a joke; then people go away tor of new forms. The generator has to rect but plausible, as well as con- thinking they enjoyed your talk. How- be able to receive information from firmed, theories proved worth consid- ever, I don’t have any good abstrac- the changing environment and make ering. The disconfirmed theories pro- tions for good jokes in AI, so this is new forms that are adapted to these vided a theory type that might be use- one piece of compiled hindsight that changes. ful in cases outside the biological won’t be instantiated here. Historically, an instructive theory examples from which the type was ab- was also proposed in immunology. stracted. Also, understanding the rea- The template theory of antibody for- son the disconfirmed theories failed in Acknowledgments mation (Pauling 1940) proposed that these biological cases provided addi- Lindley Darden gratefully acknowledges amorphous preantibodies were gener- tional hindsight about appropriate the support of the University of Maryland ated. An invader was then used as a conditions for instantiating one or the Institute for Advanced Computer Studies template and the preantibody formed other of the two types. and the College of Arts and Humanities, into an antibody by wrapping around University of Maryland, for this work, it. Neither of the instructive theories Thanks to Joseph Cain, Subbarao Kamb- was confirmed historically. In both hampati, Joel Hagen, Pamela Henson, Roy Conclusion Rada, and colleagues at the National biological evolution and antibody pro- Library of Medicine for comments on an duction, the selection theories have The compiled hindsight from the study of selection and instructive the- earlier version of this talk. Thanks to Mark proved the successful types. Hindsight Weiser, JamesPlatt, David Itkin, and Gwen thus shows, as Edelman succinctly ories is the following: find a current problem in which something is to Nelson for help in getting the manuscript said, “It is clear from both evolution- into final form. ary and immunological theory . . . that change over time to fit something else; devise good abstractions of the theory in facing an unknown future, the fun- References damental requirement for successful types for adaptation problems, namely, Burnet, F. M 1957 A Modification of adaptation is preexisting diversity” selection and instructive theories; Jerne’s Theory of Antibody Production

40 AI MAGAZINE Using the Concept of Clonal Selection. The 85-1071, Dept. of Computer Science, Stan- Polya, G. 1957. How to Solve It Garden Australian fournal of Science 20:67-69. ford Univ. City, N.Y. Doubleday. Campbell, D. T 1974 Unjustified Variation Hanson, N R 1961 (1970). Is There a Logic Popper, K. 1965 The Logic of Scientific Dis- and Selective Retention in Scientific Dis- of Scientific Discovery? In Current Issues covery. New York Harper Torchbooks covery In Studies in the Philosophy of Biol- in the Philosophy of Science, eds. H. Feigl Quine, W. 0 1969. Set Theory and Its ogy, eds. F J Ayala and T Dobzhansky, and G. Maxwell New York: Holt, Rinehart, Logic Cambridge, Mass : Harvard Univer- 139- 16 1. Berkeley: University of California and Winston. Reprint. In Readings in the sity Press Press Philosophy of Science, ed. B. Brody, 620- 637. Englewood Cliffs, N J.: Prentice-Hall Reggia, J.; Dana, N ; Wang, P.; and Peng, Y. Charniak, E, and McDermott, D. 1985. 1985. A Formal Model of Diagnostic Infer- Introduction to Artificial Intelligence Harm , R. 1970 The Principles of Scientific ence. Information Sciences 371227-285. Reading, Mass.: Addison-Wesley. Thinking. Chicago: University of Chicago Sacerdoti, E. 1974 Planning in a Hierarchy Press. Cohen, P., and Feigenbaum, E., eds. 1982. of Abstraction Spaces. Artificial Intelli- Handbook of Artificial Intelligence, vol. 3. Hesse, M. 1966. Models and Analogies in gence 5:115-135. Los Altos, Calif.: Kaufmann. Science Notre Dame, Ind.: University of Notre Dame Press. Schank, R. 1983 The Current State of AI: Darden, L. 1983 Reasoning by Analogy in One Man’s Opinion. AI Magazine 4( 1):3-8. Scientific Theory Construction. In Proceed- Kedar-Cabelli, S. 1985. Purpose-Directed Schank, R., and Abelson, R. P. 1977. Scripts, ings of the International Machine Learning Analogy In Proceedings of the Seventh Plans, Goals, and Understanding. Hills- Workshop, ed. R. Michalski, 32-40. Urbana- Annual Conference of the Cognitive Sci- dale, N.J.: Lawrence Erlbaum. Champaign, Ill : Univ. of Illinois. ence Society, 150-159. Irvine, Calif. Suppes, P. 1972. Axiomatic Set Theory. Darden, L. 1980. Theory Construction in Korf, R. E. 1985. An Analysis of Abstraction New York: Dover. Genetics. In Scientific Discovery: Case in Problem Solving. In Proceedings of the Studies, ed. T. Nickles, 151-170. Dor- 24th Annual Technical Symposium, 7-9. Thagard, P. Forthcoming. Computational drecht, Netherlands: Reidel Gaithersburg, Md : Washington, D.C., Philosophy of Science. Cambridge, Mass.: Darden, L., and Maull, N. 1977. Interfield Chapter of the Association for Computing MIT Press. Theories. Philosophy of Science 4443-64. Machinery Thagard, P., and Holyoak, K 1985. Discov- ering the Wave Theory of Sound: Inductive Darden, L., and Rada, R Forthcoming. Komfeld, W., and Hewitt, C. 1981. The Sci- Inference in the Context of Problem Solv- Hypothesis Formation via Interrelations. entific Community Metaphor IEEE Trans- ing. In Proceedings of the Ninth Interna- Analogica: The First Workshop on Analogi- actions on Systems, Man, and Cybernetics tional Joint Conference on Artificial Intelli- cal Reasoning, ed. A. Prieditis. London: SMC-11: 24-33. gence, 610-612. Menlo Park, Calif.: Ameri- Pitman. Langley, P j Bradshaw, G ; and Simon, H. can Association for Artificial Intelligence. Edelman, G, and Mountcastle, Y. 1978. 1983. Rediscovering Chemistry with the Utgoff, P. 1986 Shift of Bias for Inductive The Mindful Brain, Cortical Organization BACON System. In Machine Learning, Concept Learning. In Machine Learning, and the Group Selective Theory of Higher eds. R. Michalski, J. Carbonell, and T. vol. 2, eds R. Michalski, J. Carbonell, and Brain Function Cambridge, Mass.: MIT Mitchell, 307-329. Palo Alto, Calif.: Tioga. T. Mitchell, 107-148. Los Altos, Calif : Mor- Press. Langley, P.; Zytkow, J.; Bradshaw, G.; and gan-Kaufmann. Friedland, P. 1979. Knowledge-Based Experi- Simon, H. 1983 Three Facets of Scientific Zeigler, B., and Rada, R. 1984. Abstraction ment Design in Molecular Genetics, Tech- Discovery. In Proceedings of the Eighth in Methodology: A Framework for Comput- nical Report, 79-771, Dept. of Computer International Joint Conference on Artificial er Support Information Processing and Science, Stanford Univ. Intelligence, vol. 1, 465-468. Menlo Park, Calif.: American Association for Artificial Management 20:63-79 Friedland, P., and Kedes, L. 1985 Discover- Intelligence. Zytkow, J., and Simon, H. 1986. A Theory ing the Secrets of DNA. Communications of Historical Discovery: A Construction of of the ACM 28:1164-l 186 Langley, P.; Zytkow, J.j Simon, H.; and Brad- shaw, G. 1986. The Search for Regularity: Componential Models. Machine Learning Genesereth, M 1980. Metaphors and Mod- Four Aspects of Scientific Discovery. In 1:107-136. els. In Proceedings of the First Annual Machine Learning, vol. 2, eds. R. Michals- National Conference on Artificial Intelli- ki, J Carbonell, and T Mitchell, 425-469. gence, 208-211. Menlo Park, Calif.: Ameri- Los Altos, Calif.: Morgan-Kaufmann. can Association for Artificial Intelligence. Lenat, D. 1983. The Role of Heuristics in Gentner, D. 1983. Structure Mapping--A Learning by Discovery: Three Case Studies Theoretical Framework for Analogy. Cogni- In Machine Learning, eds. R. Michalski, J. tive Science 7:1.55-170 Carbonell, and T. Mitchell, 243-306 Palo Gick, M, and Holyoak K. 1983. Schema Alto, Calif.: Tioga. Induction and Analogical Transfer. Cogni- Lindsay, R.; Buchanan, B G.; Feigenbaum, tive Psychology 15: l-38. E. A.; and Lederberg, J. 1980. Applications Glymour, C.; Kelly, K.; and Scheines, R. of Artificial InteLligence for Organic Chem- 1983 Two Programs for Testing Hypotheses istry, The DENDRAL Project New York: of Any Logical Form. In Proceedings of the McGraw-Hill. International Machine Learning Workshop, McCorduck, P 1979. Machines Who ed. R. Michalski, 96-98. Urbana-Cham- Think San Francisco: Freeman. paign, Ill.: Univ. of Illinois. Pauling, L. 1940 A Theory of the Forma- Greiner, R. 1985 Learning by Understand- tion of Antibodies. fournal of the American ing Analogies, Technical Report, STAN-CS- Chemical Society 62:2643-2657.

SUMMER 1987 41