Scientific Theory Formation Through Analogical Inference

In the course of trying to further understand the world around him, man repeatedly attempts to find explanations for observed physical phenomena . This scenario applies to both scientists working in the laboratory and non-scientists assimilating everyday experiences . People don't carry around a full theory of the world in their heads; they make conjectures as a result of everyday experiences . Theories are tentatively proposed, they are checked to see if they adequately account for observed behavior, and sometimes experiments are performed to confirm predictions sanctioned by the new theory . One of the goals of Artificial Intelligence is to construct intelligent, autonomous systems : They too must possess the flexibility to form and refine physical theories in the course of interacting with the world . This paper presents an investigation into the process of scientific model formation ; specifically, the discovery and refinement of qualitative models of the physical world. First, general principles underlying all scientific theory formation are discussed . A theory of analogical learn ing, called Verification-Based Analogical Learning, is then presented which adheres to these basic principles . This theory shows how analogy may be used to discover and refine scientific models of the physical world through simple observation and interaction with physical phenomena. It describes how an initial model of a domain may be constructed to explain a new, inexplicable situation and how a verification is constructed to demonstrate that the new model adequately explains the observed behavior. It goes on to show how a simple planner with a knowledge of naive physics may be used to verify certain types of predictions indicated by the new theory . Examples are taken from an implemented system, which uses the theory to discover and verify qualitative models of processes such as water flow and heat flow .