Maximization doesn't work

[zStupan]

Steps to reproduce

from niapy.task import OptimizationType, Task
from niapy.algorithms.basic import BatAlgorithm

task = Task(problem='sphere', dimension=10, max_iters=250, optimization_type=OptimizationType.MAXIMIZATION)
algorithm = BatAlgorithm()
print(algorithm.run(task))

Expected behavior
Should maximize the sphere function, i. e print solution full of ± 5.12 with a fitness of 262.144 or something very close to that.

Actual behavior
None and -inf get returned, which means the stopping condition evaluates to True on the first iteration.

I'm working on a fix, should be up in a few minutes.

[firefly-cpp]

Thanks @zStupan. Please check maximization for the other algorithms too. I remember we had several problems with GWO algorithm in the past years.

[zStupan]

Most Algorithms work fine for maximization. There are, however, some issues:

• CrowdingDifferentialEvolution just doesn't work at all and doesn't have a test. I'll probably comment and add a TODO..
• CMAES, works only sometimes. It appears that it only works when the number of evals is 10,000 or below, very weird.
• KrillHerdV11, AnarchicSocietyOptimization and GlowwormSwarmOptimizationV1 all raise RuntimeWarning: invalid value encountered in double_scalars, no idea why or how to fix it. I think it's because there are inf fitness values in the population fitness.
• If the stopping condition is true before the iterations begin, for example if there's an individual in the initial population, whose fitness is already less / greater than the cutoff value, there could be bugs, because 1 iteration will run regardless.

[GregaVrbancic]

@zStupan thanks for pointing the algorithms which are not working properly. If we are unable to solve the problems with those algorithms I would suggest, we temporarily exclude them from release. Maybe leave them in a separate branch and fix them later? What do you think @firefly-cpp?

[firefly-cpp]

@GregaVrbancic, I totally agree with you. We should put all immature implementations of algorithms in separate branches. We can fix them after the release of 2.0 version of NiaPy.

[firefly-cpp]

@zStupan, I suggest that we include only one version of KrillHerd algorithm which is actually associated with original publication.

[zStupan]

Crowding DE should definitely be excluded.

I figured out what's causing the warning in GSOv1:

NiaPy/niapy/algorithms/basic/gso.py

Lines 301 to 302 in cc77c26

	def calculate_luciferin(self, luciferin, fitness):
	return np.fmax(0.0, (1 - self.rho) * luciferin + self.gamma * fitness)

The function value of the sphere function is always greater than 0, and maximization is the minimization of -f(x), so our function values will always be negative. That means luciferin will always be all zeros, which causes division by zero in the probabilities function, making some probabilities infinite. I solved this by adding epsilon to the denominator in the probabilities function.

In ASO, I fixed it by changing:

NiaPy/niapy/algorithms/other/aso.py

Lines 418 to 421 in cc77c26

	x_best, x_best_fitness = self.update_personal_best(population, population_fitness,
	np.zeros((self.population_size, task.dimension)),
	np.full(self.population_size,
	task.optimization_type.value * np.inf))

to:

x_best, x_best_fitness = self.update_personal_best(population, task.optimization_type.value * population_fitness,
                                                           np.zeros((self.population_size, task.dimension)),
                                                           np.full(self.population_size, np.inf))

Basically flipping the inequality operator if the optimization type is maximization.

I couldn't figure out KrillHerdV11 and CMAES.

[firefly-cpp]

Awesome @zStupan.

KH and CMAES should go to the separate branches. We will fix them when time allows us to do it.