Construction and Refinement of Justified Causal Models Through Variable-Level Explanation and Perception, and Experimenting
Author(s)Doyle, Richard J.
The competence being investigated is causal modelling, whereby the behavior of a physical system is understood through the creation of an explanation or description of the underlying causal relations. After developing a model of causality, I show how the causal modelling competence can arise from a combination of inductive and deductive inference employing knowledge of the general form of causal relations and of the kinds of causal mechanisms that exist in a domain. The hypotheses generated by the causal modelling system range from purely empirical to more and more strongly justified. Hypotheses are justified by explanations derived from the domain theory and by perceptions which instantiate those explanations. Hypotheses never can be proven because the domain theory is neither complete nor consistent. Causal models which turn out to be inconsistent may be repairable by increasing the resolution of explanation and/or perception. During the causal modelling process, many hypotheses may be partially justified and even leading hypotheses may have only minimal justification. An experiment design capability is proposed whereby the next observation can be deliberately arranged to distinguish several hypotheses or to make particular hypotheses more justified. Experimenting is seen as the active gathering of greater justification for fewer and fewer hypotheses.
MIT Artificial Intelligence Laboratory
MIT Artificial Intelligence Laboratory Working Papers, WP-284