FLOW OF INFORMATION THROUGH AN ARTIFICIAL NEURAL-NETWORK

The patterns of information flow through an artificial neural network are examined in terms of how and why a network characterises input/output relationships, and what insight these patterns give as to the characteristics of the network that could be changed to improve its description of a system. The prediction of vapour-liquid equilibrium in terms of bubble-point conditions is used as a case study and shows that the network is capable of identifying the intrinsic characteristics of the system. However the accuracy of the prediction depends on the region of the input/output data space considered, drawing attention to the difficulties encountered by the empirical structuring of a network. The ability to identify the type and strength of the relationships between process variables indicates that a priori knowledge of the system could be used to relate parts of the network to dominant elements of the intrinsic model. This implies there could be advantages to be gained by exploiting knowledge of the system to maximise the information content captured by the network and establish a systematic way of designing its structure.