Learning by understanding analogies (reasoning)
Russell Greiner · 1985
The phenomenon of learning has intrigued scholars for ages; this fascination is reflected in Artificial Intelligence, which has always considered learning to be one of its major challenges. This dissertation provides a formal account of one mode of learning, learning by analogy. In particular, it defines the useful analogical inference process (UAI), which uses a given analogical hint of the form A is like and a particular target problem to map known facts about B onto proposed conjectures about A. UAI only considers conjectures which are useful to the target problem; that is, the conjectures must lead to a plausible solution to that problem. To construct a procedure which can effectively find these useful analogies, we use two sets of heuristics to refine the general UAI process. The first set is based on the intuition that useful analogies often correspond to coherent clusters of facts. This suggests that UAI seeks only the analogies which correspond to common abstractions, where abstraction are relations which encode solution methods to past problems. The other set of rules embody the claim that better analogies impose fewer constraints on the world. Basically, these rules prefer the analogies which require the fewest additional conjectures. This dissertation also describes a running program, NLAG, which implements this model of analogy. It is then used in a battery of tests, designed to empirically validate our claim that UAI is an effective technique for acquiring new facts. This data also demonstrates that the heuristics are effective, and suggests why. In summary, the primary contributions of this research are (1) a formal definition of UAI, described semantically (using a new variant of Tarskian semantics), syntactically and operationally; (2) a collection of heuristics which efficiently guide this process towards useful analogies; and (3) various empirical results, which illustrate the source of power underlying this approach.