An interactive scientific theory revision system
Scientists are often presented with vast amounts of laboratory/observational data which need interpretation. These data can often be generalised to produce theories that can be subsequently used for the classification of unseen data. The construction of theories is a hard task which becomes even harder when new data reveals errors and gaps in the theory, thus requiring appropriate corrections. One case in point is the theory for predicting NMR spectroscopy coupling constants for sugar molecules. Revising this theory is difficult because it requires both NMR and Chemistry expertise, as well as heuristic knowledge of the interpretation of spectra. Moreover, as there are relatively little data, statistical-based analysis cannot be used effectively. Taking these factors into account, we developed a scientific theory revision system, CRITON, to assist NMR domain experts in the correction & maintenance of this theory. The major characteristics of the system include the following: First, the ability to enhance the domain's Concept Description Language (CDL) allows the system to search efficiently in a restricted hypothesis space and automatically adjust its CDL only when the latter is shown to be insufficient, i.e. when it cannot discriminate between two training examples of different classes. Second, the possibility of heuristic knowledge in the form of domain-specific biases which can be customised to the current application; and third, user selection of theory modifications proposed by the system. This characteristic along with the incremental processing of training examples suggests the use of CRITON as an interactive theory exploration tool, a scientific aid. One special aspect of the system is the knowledge representation language which is a hybrid between production rules and a graph representation. This hybrid representation allows CRITON to revise relational theories (e.g. the coupling constant theory) which cannot be succinctly represented as production rules. Although originally CRITON was designed for use in the NMR spectroscopy domain it has evolved to become a domain-independent program; a series of experiments have been run comparing CRITON's performance with those of FOCL and GOLEM. This has led to the claims that CRITON is a general relational learning system.