Toward context dependent classification of infra-red spectra by energy minimization

An architecture for reimplementing an existing expert system is described. The original application, called GOLD, identifies minerals in rock samples, using infrared spectral data, and was developed in the Prospector formalism. The development of GOLD´s knowledge base became cumbersome when it was necessary to describe how the spectrum of one mineral is affected by the presence of another mineral in the same sample. This is due to Prospector´s one-way flow of information in the inference net, from data to goals. Any form of unidirectional knowledge representation would suffer the same problem. A bidirectional energy minimization network is a more suitable form of knowledge representation. This is illustrated on a simplified version of the spectral interpretation problem