Patricia.Mackiewicz@isl1.ri.cmu.EDU (04/17/87)
TOPIC: Reactive Learning: Experimentation and Decompilation SPEAKER: Jaime Carbonell, CMU WHEN: Tuesday, April 21, 1987, 3:30 p.m. WHERE: Wean Hall 5409 Most symbolic learning approaches have been purely empirical (inductive) or purely analytical. The former extracts a general concept from a set of empirical observations, whereas the latter composes primitive concepts into larger units (chunks, macro-operators, "explanations", etc.). Analytical methods include explanation-based learning, capable of exploiting a complete domain theory to learn complex concepts from very few instances. However, the domain theory may be partial, and judicious integration of empirical and analytical methods may prove far superior to either method alone. Reactive experimentation is a case in point: partial domain knowledge is used to formulate hypotheses, and empirical data from the experiments is used to formulate new concepts or modify existing ones. Decompilation maps complex empirical observations into comprehensible operational units using analytical techniques. Both methods for combining analytical and empirical approaches are explored with the objective of creating robust learning systems.