Population level stagnation

[seallard]

If no fitness improvement has been observed in the entire population over a certain threshold, the original NEAT implementation replaces the entire population with offspring from the two best species.

[seallard]

Here is one way to do it:

    def stagnation_check(self):

        for species in self.species:

            if species.leader.original_fitness > self.champion_fitness:
                self.champion_fitness = species.leader.original_fitness
                self.epochs_since_improvement = 0

        self.epochs_since_improvement += 1

        if self.epochs_since_improvement > self.config.population_maximum_stagnation:
            self.stagnation_reset()

    def stagnation_reset(self):
        """Replace the population with offspring from the two best species. """

        if len(self.species) > 1:
            print("RESETTING POPULATION!")

            self.species[0].expected_offspring = floor(self.config.population_size/2)
            self.species[1].expected_offspring = floor(self.config.population_size/2)

            for species in self.species[2:]:
                species.expected_offspring = 0

            self.champion_fitness = 0
            self.epochs_since_improvement = 0

[seallard]

But the performance seems to get worse. Probably better to just rely on the speciation and mutations. Seems like a weird thing to do. Might be for "premature convergence prevention". https://en.wikipedia.org/wiki/Premature_convergence