Hands-On Neuroevolution with Python

Hyperparameter selection

Compared to the previous experiment described in this chapter, double-pole balancing is much harder to solve due to its complex motion dynamics. Thus, the search space for a successful control strategy is broader and requires a more diverse population. To increase the diversity of the population, we increase its size to become 10 times bigger than with a single-pole balancing experiment.

The fitness termination threshold remains the same as shown here:

[NEAT]fitness_criterion = maxfitness_threshold = 1.0pop_size = 1000reset_on_extinction = False

To intensify evolutionary diversity further, we increase the probabilities of adding new nodes and connections, as well as changing the configuration scheme of initial connections. ...