pecific examples of robot patterning arethedevelopmenttechniquesof"showandtell"robot and the development control (seesection6.3).While commercialin dustrial robots have long employed patterning methods ,these methodsareusedonlyin arudimentary formandfurther applicationsdevelopmentforthemto technologyisneeded becomeusefulto NASA.Theshowandtellmodeofrobot actionhasapparentlyandinvestigated notbeenidentifiedto anylargedegree,andseemswhere to beanareaNASA shouldtakean immediateandlargeinterestbothin the theoryandtheapplications aspects.
6.2 Learning and Hypothesis Formation The Titan exploration mission description, documented in chapter 3, discusses the characteristics of a machine intelligence system possessing autonomous self-learning. This capability, its relation to state-of-the-art AI, and the new research directions it demands are summarized below. 6.2. 1 Characteristics For a machine to learn a previously unknown environment involves both the deployment of knowledge structures correct for known environments and the invention (or discovery) of new knowledge structures. A machine intelligence system which learns could formulate (1) hypotheses which apply existing concepts, laws, theories, generalizations, classification schemes, and principles to the events and processes of the new environment, and (2) hypotheses which state new concepts, laws, and theories whenever the existing ones are inadequate. Different logical patterns of inference underlie the formation of these types of hypotheses. Analytic inferences support the formation of hypotheses which apply existing concepts, laws, and theories. Inductive and abductive inferences support the invention of hypotheses which state new concepts. Analytic, inductive, and abductive inference are mutually and logically distinct -one of them cannot be replaced by some combination of the others (see section 3.3.3 and compare Fann, 1970; Hanson, 1958; Lakatos, 1970a, 1970b, 1976; Peirce, 1960, 1966).
6.2.2 State-of-the-art in AI State-of-the-art AI lacks adequate and complete treatments of all three inferential classes necessary for the development of machine intelligence systems able to learn in new environments. Analytic inferences receive the most complete treatment. For instance, rule-based expert systems can apply detailed diagnostic classification schemes to data on events and processes in some given domain and produce appropriate identifications (Buchanan and Lederberg, 1971; Duda et al., 1978; Feigenbaum, 1977; Martin and Fateman, 1971; Pople, 1977; Shortliffe, 1976). An expert system such as PROSPECTOR can identify a restricted range of ore types and map the most likely boundaries of the deposit when given survey data about possible ore sites (Duda et al., 1978). However, these systems consist solely of complicated diagnostic rules describing the phenomena in some domain. They do not include models of the underlying physical processes. In general, state-of-the-art AI treatments of analytic inference fail to link the detailed classification schemes used in these inferences with the fundamental models required to deploy this detailed knowledge with maximal efficiency. Inductive inferences receive a less complete treatment although some significant advances have been made. For example, Hajek and a group of co-workers at the Czechoslovak Academy of Sciences have, over the past 15 years, developed and implemented systems of mechanized inductive generalization (Hajek and Havranek, 1978). They do not take the approach of "inverse deduction" which has been explored by Morgan (1971,1973). Instead, the Czech group has developed techniques for moving from data about a restricted number of members of a domain, to observation statement(s) which summarize the main features or trends of these data, to a theoretical statement which asserts that an abstractive feature or mathematical function holds for all members of the domain. (For instance, see table 6.6.) Though they allow a role for what they call "theoretical assumptions," in moving from observation to theoretical statements they have concentrated their work on formulating the rational inductive inference rules for bridging the gap between the two – TABLE 6.6.-SAMPLE INFERENCE DATA
Weight, Weight of rat
Rat no.
g kidney, mg
362 1432 373 1601 376 1436 407 1633 411 2262
Observa-Therefore, the observed tion weights of the kidneys statement have the same order as the weights of the rats with one exception. Theo-Therefore, the weight of a retical rat's kidney is positively statement dependent on the weight of the rat. thoughit isnotclearthattheirsystemcapturesfu
