Tabletop: An emergent, stochastic computer model of analogy-making.

Abstract

The key notion underlying the research presented in this dissertation is my conviction that the cognitive mechanisms giving rise to human analogy-making form the very basis of intelligence. Our ability to perceive and create analogies is made possible by the same mechanisms that drive our ability to categorize, to generalize, and to compare different situations. A theory of concepts and analogy-making, and a computer model of that theory--"Tabletop"--are presented. The computer model operates in a microworld that was carefully designed to be rich enough to allow the mechanisms posited by the theory to be put to the test, yet not so complex that the program would be overwhelmed by irrelevant details typical of real-world situations. This work can be considered midway between traditional symbolic and the more recent connectionist approaches to the modeling of intelligence. The architecture of the program relies on a context-sensitive concept network modeling long-term memory, in which concepts gain, lose, and spread activation. This network continually interacts with a workspace modeling short-term memory, in which structures are dynamically built, modified and destroyed. All of the work of the program is done by a large number of competing, semi-independent, local agents. Nondeterminism in the form of biased probabilistic decision-making permeates every level of the program. Nonetheless, robust coherent behavior, in the form of high-level representation-building and analogy-making, emerges as a statistical result of the myriad decisions made during processing. The details of Tabletop's behavior are shown by watching the program's progress as it solves a single problem. In addition, the program's "personality" is revealed by statistics summarizing key features of large numbers of runs on several families of systematically related analogy problems.