Skip to main content

AN ARTIFICIAL INTELLIGENCE SYSTEM FOR THEORY-DRIVEN DISCOVERY OF SCIENTIFIC LAWS

Objective


The problem considered was the prediction of torsional angles in molecules whose topology is known, given certain information which can be extracted by an expert from the molecules' nuclear magnetic resonance (NMR) spectra. Traditional knowledge acquisition techniques were used to obtain an approximation of the problem solving knowledge. A knowledge base refinement system was next designed and implemented to improve the initial theory; the system is provided with a series of examples for each of the classes, where each example is described as a vector of attributes and associated values. The major elements of the system are: incremental processing of the training examples of the specified classes; specializations and generalizations of class concepts; detection of inconsistencies in the empirical theory and the training examples; resolution of the inconsistency using one (or more) of a known set of chemical descriptors which the expert had originally judged to be irrelevant to the problem domain.

HUME is an integrated discovery system which implements a number of different theory driven learning strategies, both deductive and inductive. It is also intended to allow the integration of data driven inductive discovery systems, such as ARC. Given one from a set of observations, HUME finds an explanation for this observation in terms of the theories and knowledge available to the system, such that this explanation can be generalized to explain a significant subset of all observations. However, since the system will in most cases by dealing with incomplete domain theories, a significant subpart of this task will involve theory and law construction, by the use of a theory construction algorithm and the application of informal qualitative models (IQMs). At its most basic level, HUME is a backward chaining Horn clause theorem prover. HUME has available to it a set of other inference modes which it can use in an attempt to construct the missing parts of the domain theory. These other inference modes are a combination of abduction and forward deduction, together with the inductive step of the DISCIPLE algorithm for theory construction.
We propose to build a machine discovery system that will use an incomplete theory of a scientifc domain to guide its search for empirical laws and exceptions to them, and will use there to refine its therory. We plan to apply the system to problems in plysical chemistry. The system will comprise five interacting modules, corresponding to separate subproblems and types of scientific reasoning (1) The theory Operationalisation Module will use the domain theory to predict the form of the laws to be discovered. (2) The law Discovery Module will use this information to guide the determination of the exact forms of the laws, using statistical techniques. (3) The data Evaluation Module will determine whether points that do not fit the laws are errors of special cases, and identify possible reasons for the deviations from predictions. (4) The Law Evaluation Module will evaluate and compare alternative candidate laws. (5) The Theory Revision Module will use the empirical laws and the anomalous cases discovered to refine the domain theory.

Funding Scheme

CSC - Cost-sharing contracts

Coordinator

University of Aberdeen
Address
Regent Walk
AB9 1FX Aberdeen
United Kingdom

Participants (1)

CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE
France
Address
Bètiment 210 - Université Paris-sud
91405 Gometz La Ville