Genetic algorithm /w Neural Network playing snake is not improving

I noticed that in your pseudocode, upon creating each new generation, the parent generation is completely wiped out and only the child generation is retained. This can naturally lead to lowering fitness levels, since there is nothing guaranteeing that the offspring will have fitness levels comparable to those in the parent generation. In order to ensure that the fitness levels are non-decreasing, you must either merge the parent and child generations and prune the weakest members (which I would recommend), or you can require that the offspring-generating function produces offspring at least as fit as the parents (by many trials and errors).

If you decide to focus on the offspring-generator, one way to (somewhat) guarantee improved offspring is to implement asexual reproduction by way of simply adding a small amount of noise to each weight vector. If the noise level is small enough, you can generate improved offspring with a success rate of up to 50%. Bigger noise levels, however, allow faster improvement and they help to jump out of local optima, even if they have success rates lower than 50%.

You are only mutating 5% of the population, not 5% of the "genome". This means your population will be fixed incredibly quickly - https://en.wikipedia.org/wiki/Fixation_(population_genetics) .

This makes sense why the population isn't doing very well, because you are only exploring a very small area of the fitness landscape ( https://en.wikipedia.org/wiki/Fitness_landscape ) .

You should change the mutate function to mutate 5% of the genome (ie weights between nodes). Feel free to play around with the mutation rate as well - different problems perform better at different mutational rates.

If you are worried about loosing the current "best genome", a typical approach in evolutionary computation is to copy the individual with the highest fitness over to the next generation without mutation.

(Sorry, this probably should have been a comment but I don't have enough reputation).

I had this exact problem for a week and I only just figured out what the issue is. When assigning brains from one snake to another, make sure to use array.copy. E.g do this:

a = np.array([1, 2, 3, 4])b = a.copy()

and not this:

a = np.array([1, 2, 3, 4])b = a

This is because python will sometimes make b and a share the same memory in the second case, and so whenever you reassign b, a will also be reassigned.