GeneticSnake | Genetic Algorithm to teach a Neural Network | Machine Learning library

I'm not surprised your snakes are dying.

Let's take a step back. What is AI exactly? Well, it's a search problem. We're searching through some parameter space to find the set of parameters that solve snake given the current state of the game. You can imagine a space of parameters that has a global minimum: the best possible snake, the snake that makes the fewest mistakes.

All learning algorithms start at some point in this parameters space and attempt to find that global maximum over time. First, let's think about MLPs. MLPs learn by trying a set of weights, computing a loss function, and then taking a step in the direction that would further minimize the loss (gradient descent). It's fairly obvious that an MLP will find a minimum, but whether it can find a good enough minimum is a question and there are a lot of training techniques that exist to improve that chance.

Genetic algorithms, on the other hand, have very poor convergence characteristics. First, let's stop calling these genetic algorithms. Let's call these smorgasbord algorithms instead. A smorgasbord algorithm takes two sets of parameters from two parents, jumbles them, and then yields a new smorgasbord. What makes you think this would be a better smorgasbord than either of the two? What are you minimizing here? How do you know it's approaching anything better? If you attach a loss function, how do you know you're in a space that can actually be minimized?

The point I'm trying to make is that genetic algorithms are unprincipled, unlike nature. Nature does not just put codons in a blender to make a new strand of DNA, but that's exactly what genetic algorithms do. There are techniques to add some time of hill climbing, but still genetic algorithms have tons of problems.

Point is, don't get swept up in the name. Genetic algorithms are simply smorgasbord algorithms. My view is that your approach doesn't work because GAs have no guarantees of converging after infinite iterations and MLPs have no guarantees of converging to a good global minimum.

What to do? Well, a better approach would be to use a learning paradigm that fits your problem. That better approach would be to use reinforcement learning. There's a very good course on Udacity from Georgia Tech on the subject.