Learning retrodictive knowledge from scientific laws : the case of chemical kinetics

We consider the problem of extracting retrodictive knowledge from scientific laws used ordinarily only to predict. In particular, a method is developed which synthesizes rules of experiment-interpretation from the basic law of chemical kinetics. Previous work in AI on transforming predictive knowledge into convenient retrodictive knowledge has been within the subfield of diagnosis. The current work extends the idea to the domain of elucidation of causal mechanism. Refutation rules are synthesized by discovering invariants within a parameterized system of equations. The choice of invariants to look for is guided by four criteria. A principle of stable refutation, based on the character of experimental data, is derived from the non-rescindible nature of refutation. Three other criteria contribute to the practicality, generality, and reliability of the rules. The invariants chosen are tested by systematic sampling of a system parameter-space. Hence, the rules, which check that an invariant holds for experimental data, are established by induction from simulation data. The synthesized rules serve in practice as reliable disconfirmatory evidence, rather than refutations, due to their inductive origin as well as to the uncertainty of experimental data. The rules will be applied within the context of ongoing work on elucidation of chemical-reaction networks. The author is supported by a Graduate Fellowship from the Avionics Laboratory at Wright-Patterson Air Force Base, awarded by Universal Energy Systems. The views and conclusions contained m this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. government or Universal Energy Systems.