CHAPTER 5
 SIMULATIONS WITH REAL DATA SET

 In this chapter we present the results of segmentation on the original data set with

the different models and some parameter tuning, and we analyze the system in the light of

all these results.

 5.1 Segmentation Results

 From the results obtained previously, approach IV is the one that showed the most

promise, providing that the gate implements harder competition, which is done by

increasing the competition parameter to M=5, except for the MLP experts for which M=2

appears to be sufficient. Also, with the real non-stationary data, at first simulation showed

that the memory depth might be too long, so for all simulations below, for ys =0.02 is

used and helps in shortening the delays before a switch is detected.

5.1.1 Linear Model

 Figure 5-1 shows simulation with linear experts, which quite expectedly does not

perform better than on the test set, and still has a high prediction error.