Graphical representation of model performance as an aid for input selection in greenhouse models
The selection of input variables is the most sensitive stage of model building, involving technical skills as well as expert knowledge of the problem to be modelled. One way to improve the model fit is to provide it with more input information, i.e., more input signals. However, depending upon the modelRSQUOs type and the specific algorithms involved, adding a large number of inputs can also lead to a bad model fit. An additional reason to keep the model simple is the computational cost, both during training and when running simulations. It also implies a higher cost in sensors and equipment when inputs are expected to be measured. This work proposes a new type of graphical method which can be an aid to visualize the relative contribution of a single input to the model. The diagram consists of two reference lines, representing the simplest (less input variables) and most complex (all available inputs) models. A number of points represent the model fit of the models generated by the addition (in the first case) and subtraction (in the second) of a single input variable. In several examples, the model fit is calculated as the MSE after 500 training epochs of an artificial neural network. The examples include 3 different models inside a tomato greenhouse: Climate prediction, crop transpiration rate and crop photosynthesis rate. The distances from the individual points to the two reference lines give a clear hint on which input signals provide the best contribution to model fit and should be considered to be included in the final model. Although these examples use the MSE as measure of model fit, the proposed graphical procedure allows comparing models using any quantitative measure.
Miranda, L., López-Cruz, I.L., Lara, B. and Schmidt, U. (2017). Graphical representation of model performance as an aid for input selection in greenhouse models. Acta Hortic. 1182, 227-234
model performance, MSE, plot, model comparison, greenhouse models, artificial neural networks