Learning in Deduction by Knowledge Migration and Shadowing

A method of deductive learning is developed to control deductive inference Our goal is to im prove problem solving time by experience when that experience monotonically adds knowledge to the knowledge base In particular for deductive reasoning systems where partial results are saved dur ing a derivation and at least some partial results are themselves deduction rules we suggest ways of taking maximal advantage of old partial results and avoiding the regeneration of partial results when solving new problems We approach the problem of accumulating past deductive experience and using it in subsequent similar deductions by the scheme of knowledge migration and knowledge shadowing Knowledge migration generates a speci c rule from a general rule during a deduction and accumulates deduction experience that is represented by the speci city relationship between the general migrat ing rule and the speci c migrated rule Knowledge shadowing uses deduction experience obtained in previous inferences for faster reasoning in future inference If both general and speci c knowledge are applicable in subsequent deduction the knowledge shadowing scheme is activated to select only the speci c knowledge Three principles for knowledge shadowing are presented A preliminary result shows that knowledge migration and shadowing greatly contribute to the system performance