Learning causal relationships: an integration of empirical and explanation-based learning methods
Michael J. Pazzani · 1988
A theory of learning to predict and explain the outcomes of events is proposed. The theory integrates three sources of information: (1) Inter-example relationships: Regularities among a number of examples that reveal the conditions under which a cause produces an effect. (2) Intra-example relationships: Temporal and spatial relationships between a cause and an effect which constrain the search for a causal hypothesis. (3) Prior causal and social knowledge: Prior knowledge which predicts and explains regularities in events. I focus on the strengths and weaknesses of each source of information and their associated learning methods. I describe how the learning methods can be integrated in a complementary fashion. The proposed theory of learning is realized by a computer program that I constructed called sc OCCAM. sc OCCAM acquires causal and social knowledge by empirical techniques by exploiting inter-example and intra-example relationships. An explanation-based learning component of sc OCCAM takes advantage of prior knowledge to constrain the learning process. sc OCCAM is unique among explanation-based learning systems in that it has the ability to acquire, with empirical techniques, the background knowledge needed for explanation-based learning. For example, sc OCCAM learns about kidnapping by applying social knowledge acquired by empirical learning. The result of applying sc OCCAM to a collection of economic sanction incidents is also discussed.