Three selection techniques were used in an implementation of the genetic programming paradigm to solve the cart-pole problem. The genetic programming paradigm has been shown to successfully solve the ‘benchmark’ version cart-pole problem. Each of the selection techniques managed to direct evolution toward suitable controllers for pole balancing in this difficult problem domain. Overall, the genetic programming paradigm evolved 13 controllers, in the course of 1,500 runs, to balance the pole for 1,000 seconds. The best two controllers had RMS amplitudes of 5.4 and 7.4 centimetres from the centre of the track. The final RMS states of these two best controllers also compared favourably with the findings of previous research.

For the off-line performance of each of the selection techniques, the difference was shown to be insignificant. Thus, no confident conclusion can be made on the relative effect of selection techniques on off-line performance. However, the results of the analysis did show a significant difference in the distribution of mean fitness of the populations. This result, particularly as it conforms with the on-line performance measure found, suggests that significantly better on-line performance can be expected from the stochastic selection methods tested (expected value model and roulette wheel) over the competitive selection strategy employed by tournament selection.