Fig. 1. GENETICA searches for the best input and output combination.
As GENETICA Net Builder breeds the network population, it also searches through the input data window and the output target window for the best combination of input data and forecast horizon. This procedure is illustrated in Fig. 1.
In this illustration, the input data file has 25 columns of data. The first column is the date, the second column is the target, and the remaining 23 columns are the input indicators used to train the network populations. The input window has a window size of 10 rows with a total of 10x23 = 230 pieces of input data items, and the output forecast horizon range was set to 1 to 6 steps ahead.
The best network colored in red shows that it only requires some of the input data, and that columns number 4, 6, 11, 12, 17, 18, 21, 22, 24 are not found to be useful. For the best network, the total number of input data items has been significantly reduced from 230 to 21.
This unique feature of GENETICA Net Builder allows users to start off with a large number of input data columns, and subsequently remove the unwanted ones from the data file. Not only the input dimension can be reduced, the resulting population will learn more efficiently as it uses only those input data that were found to be useful.
Copyright © 1993-96, NIBS Pte Ltd.