AI Magazine Volume 12 Number 3 (1991) (© AAAI) Letters q Editor: system-centered research. First, the the other does. No theorist is going Thank you for the opportunity to split between the neats and scruffies to spend his or her time attempting respond to the letters by Jim Hendler, is old and institutionalized, as to bring precision to a mess of hacks, James Herbsleb and Mike Wellman Hendler points out. Few researchers kludges and “knowledge,” and no regarding my survey of the Eighth are trained in both camps. Second, system builder is apt to find the National Conference on Artificial the pathologies that researchers in attempt informative. MAD does not Intelligence (AI Magazine, Volume 12, AAAI-90 themselves attributed to mean business as usual with occa- No. 1). The letters raise many inter- focusing on just one aspect did, in sional collaborative meetings. esting points that can be roughly fact, arise. Third, the expected advan- hfAn means assessing environmen- classified as questioning the validity tages of merging systems and models tal factors that affect behavior; of the survey and questioning the did, in fact, materialize (although the modelling the causal relationships proposed MAD methodology. Mike sample was very small). Wellman between a system’s design, its envi- Wellman says, “As Cohen acknowl- says that just because I didn’t see ronment, and its behavior; designing edges, a serious problem with the system-centered and model-centered or redesigning a system (or part of a survey is that the AAAI conference research reported together, doesn’t system); predicting how the system proceedings do not accurately repre- mean it wasn’t there; I say that I will behave; running experiments to sent the field.” Actually, I did not found pathoIogies (e.g., see my sec- test the predictions; explaining unex- acknowledge that AAAI is not repre- tions Models Without Systems and pected results and modifying modeis sentative; I just raised the possibility. Systems Without Models) that and system design; and generalizing Wellman evidently believes it and, strongly suggest it wasn’t there. the models to classes of systems, further, attributes it to “the general The issue is important because I environments and behaviors. The cri- constraints of the conference forum base the MAD methodology on the terion of success is the ability to pre- and the length of proceedings assertion that system-centered dict behavior at some level of papers.” But surely Wellman would research needs models and model- accuracy useful to designers. These agree that it requires fewer than six centered research needs systems. Nei- are the basic activities of any engi- pages to say what one learned or ther Wellman nor Hendler disagreed neering science, and they will be tried to learn in one’s research? Only with this assertion but both respond- accomplished not by collaborations 17% of the papers in AAAI did so. ed vigorously to what they perceived of researchers with completely differ- The reason AI researchers say little in to be the “folly” of MAD: “Indeed, it is ent ideas of what is interesting, what six pages isn’t the page limit (note not clear to me that it would be a is evidence, what is a result, and so that scientists get by with just a few desirable end for all or even many AI on, but by indi~iiduals who want to column inches in journals such as researchers to be performing both do engineering science. Science and N&lre), it is that they are types of research” (Wellman). And, In response to James Herbsleb, let telling us the wrong stuff. To para- “If Cohen is taken (overly) seriously, me say that MAD is voluntary, as is phrase David Etherington, AAAI then every major AI researcher will the methodology he describes as an papers tell us what was done, not have to... be conversant with theo- alternative. His gibe notwithstanding what was learned. Six pages may be retical tools (math, logic, etc.), exper- I present MAD with no more hubris too little space for the former but it is imental methods (testing theory, than he and his colleagues present ample for the latter. statistics, etc.), and implementation SOAR. That is, I am enthusiastic about Wellman questions my principal strategies (blackboard architectures, MAD because it helps me answer the result that AI research is either real-time systems, etc.)-a pretty tall research questions that interest me, model-centered or system-centered order” (Hendler). Well, we have tried and I hope, as researchers in the SOAR with almost no research in the inter- dividing AI research among the theo- project hope for their methodoIogy, section He says that much more rists and the system builders, and my that MAD will prove generally useful. work is in the intersection but that, survey suggests this division does due to space limitations, only one more harm than good. Hendler’s “tall Paul R. Cohen aspect of a research project can be order” is not actually so tall, especial- Department of Computer Science reported in AAAI. I found that only ly if you think, as I do, that it could University of Massachusetts at eight papers of 150 reported both revitalize our research. Wellman and Amherst model-centered and system-centered Hendler both call for collaborative aspects. I have three reasons for efforts between model-centered and believing that the other 142 papers system-centered researchers, but this do not, in fact, represent work in the hasn’t worked, and it won’t, because intersection of model-centered and neither group cares much about what 10 AI MAGAZINE q Editor: “He’s schizophrenic”-it’s a description, The article by David West and Larry not an explanation. Analogy is a par- Travis in the Spring 1991 issue will ticular format for laying out semantic no doubt stimulate some healthy relations. As such, it was intended to debate about the status and use of describe precisely the kinds of seman- tic relations that can be involved in the mind/computer metaphor in AI. The authors relied heavily on the metaphor. Hence, analogy can be analysis of metaphor proffered by used in post-hoc descriptions of Earl MacCormic, and his distinction particular semantic relations (e.g., between epiphor and diaphor is similarity of features) which an inter- useful in the analysis of metaphor in preter has abstracted from a metaphor. science. Readers of AI Magazine But analogy is not the egg, and AAAI-91 should be apprised, however, of metaphor is not the chicken. other works on metaphor which may Robert R. Homan help clarify some of the issues. Department of Psychology Proceedings of the Ninth For one thing, there exists a con- Adelphi University National Conference on siderable literature of experimental research on metaphor comprehen- Artificial intelligence sion which bears on the claim that AmericanAssociation for metaphors can be described as simi- S Editor: Artificial Intelligence, larity statements. Suffice it to say that the scaling of the fruitfulness of Could, as his article reprinted in the July, 1991. Anaheim, California metaphors by means of simple metrics summer 1991 issue states, Marvin of the features that are shared by the Minsky really wonder what caused The theme of the 1991 conference terms in a metaphor is not so straight- the misconception that neural net- was interaction and growth. forward as West and Travis imply works and AI are conflicting activities? Furthermore, the distinction If so, he hasn’t read the many attribu- between diaphor and epiphor is tions of the conflict to his and Sey- Contents include: rather slippery, and the slipperiness mour Papert’s 1969 book Perceptrons: l CASE-BasedReasoning is compounded by West and Travis’ An Introduction to Computational additional concept of “paraphor.” Geometry l Communicationand Some traction on these distinctions Because of that book’s influence, Cooperation may be found in the work of linguist say Maureen Caudill and Charles l Constraint Reasoningand George Lakoff, and especially his dis- Butler in Naturally Intelligent Systems ComponentTechnologies cussion of the differences between (p. 171, MIT Press, 1990), “neural metaphors and “metaphor themes.” network research and development l Formal Methods in Knowledge Illumination can also be found in was brought to a near-standstill for Representation philosopher Stephen Pepper’s discus- almost two decades.” l Learning sion of the mechanist, formist, and Papert knows where the blameful contextualist metatheories. fingers are pointed: In “One AI or l Planning, Perception,and Many of these concepts and issues Many?” (Daedalus, Vol. 117, No. 1, Robotics can be found in the pages of the new Winter 1988; also The Artificial Intelli- l ReasoningAbout Physical journal, Metaphor and Symbolic Activity. gence Debate, MIT Press, 1988), he Systems Apart from overlooking the large, uses the metaphor of natural and relevant corpus of work on metaphor unnatural sisters. In the early sixties, l TractableInference and on metaphor in science, the he says (p. 3), “The artificial sister l Invited Talks West and Travis article is refreshing grew jealous and was determined to in that it does not assume the tired keep for herself the access to Lord References,index, 2 vols, old positivist stance which asserts, DARPA’S research funds. The natural approx. 1,000 pages basically, that “if it’s metaphor, and if sister would have to be slain. The it’s science, then it’s bad science.” bloody work was attempted by two $75.00 Metaphor analysis can be a useful, if staunch followers of the artificial ISBN O-262-51059-6 not essential, tool in the analysis of sister, Marvin Minsky and Seymour scientific reasoning (and reasoning in Papert... .” Papert admits to formerly Publishedby the general), and the West and Travis feeling “some hostility” (pp.