7.3.1 Anticipation with the multi-layer perceptron

For the multi-layer perceptron, the average error on the 138 test pattern series is shown in figure 7.9. The mean square error was below 2.4 pixels squared for all tested chain lengths. Starting from the second to the sixth step the square error per sector increases linearly. This linear increase matches the theoretical prediction (appendix C.3).

Figure 7.9: Anticipation performance of the MLP on the 138 test series. The mean square error (in pixels squared) per sector is shown as a function of the number l of chain links. A line is fitted to the points from link number two to six.