IMITATION OF A PROCEDURAL GREENHOUSE MODEL WITH AN ARTIFICIAL NEURAL-NETWORK

Our overall objective is to replace procedural models with neural netw orks for some reasoning activities in cognitive systems. We have initi ally attempted to imitate a procedural thermal exchange model of a gre enhouse with a number of neural networks, each of which was subjected to various amounts of learning. An evaluation method was developed wit h which the performance of each network was compared to that of the pr ocedural model. The efficacy of the evaluation method was assessed in comparison to human visual judgment. Each network was also tested for its ability to respond meaningfully to data sets which were different from its learning set. The evaluation method was found to agree with t he general trend of human visual judgment and can be used to monitor a network's progress in learning. The networks were given input values for date, time, solar radiation, exterior temperature, relative humidi ty, and wind speed, as well as one-hour lag values of the radiation an d exterior temperature. After 100,000 learning cycles, the networks ad equately mimicked the greenhouse procedural model with regard to the t hree output variables of interest: interior temperature, heating load, and ventilation. This network configuration (8 inputs, 3 outputs, 100 ,000 learning cycles) also performed acceptably when the input data us ed during recall were different from those used for learning.