The most obvious aspect of the graphs in this section is the change in SMART operator performance (due to evolution) as the generations pass. However, the main population is also changing as these generations pass, and this causes much of the changes in the graphs. While these changes (the general dynamics of the main population at generation Y) are hard to show on the graph, the relevant aspect for this chapter can largely be understood by studying the random operator curves. Since the random recombination operator remains unchanged in its operation as the generations pass, the changes to its performance curve are entirely attributable to the changing main population dynamics. Keeping this in mind will help in deciphering this section's graphed results.
Experiment-specific details will be given with the experimental results, but several pieces of experimental methodology were fixed for this entire chapter. Figures 3.4, 3.5, and 3.6 in sections 3.5.2, 3.5.3, and 3.5.4 all show data taken from 5 runs from each of 3 domains: The three domains are the sound domain mentioned in section 3.5.2 [Teller 1995b], a robotic image object recognition task [Teller and Veloso 1995a], and a difficult manufactured image classification problem [Teller and Veloso 1995d].
For the main population, all experiments used 85% recombination (SMART or RANDOM), 5% random mutation, and 10% simple reproduction. The SMART operator population was subjected to 40% random recombination, 10% random mutation (see section 3.6), and 50% simple reproduction. The main population consisted of 2800 programs. Each one was allowed one ADF and access to any of the 150 Library programs (public ADF's). The node size limit on all programs (MAIN, ADF, and library) in both the main and SMART operator populations was set to 256. The indexed memory size was set to 256 elements for both the main and SMART operator programs. The time-threshold for all programs (main population and SMART operators) was approximately equal to 8000 node evaluations. A typical run to generation 100 in PADO took about a day on a DECstation 5000/20.
The length of this chapter does not provide for a detailed study of the SMART mutation as well as the SMART recombination operators. The reason that we concentrate on the SMART recombination operators is that mutation in GP is primarily used to maintain diversity. Random mutation does this job quite well. However, discussion on SMART mutation operators can be found in [Teller 1996].