machine-learninggenetic-algorithmneat

NEAT: how does crossover occur for species with only one member

So, I'm trying to implement the NEAT(Neuroevolution of augmenting topologies) algorithm and have stumbled into a problem. How are networks in species with only one member crossed over?

One solution I came up with is to perform inter-species crossover. But I don't know if it would be effective.

Solution

  • In NEAT, there are four ways in which you can create candidate individuals for the next generation:

    1. Pass an exact copy of an individual
    2. Pass a mutated copy of an individual
    3. Do crossover using two individuals from a given species
    4. Do crossover with two individuals of different species (iter-species)

    Of course, you can always do (1). This is often applied to "elites", which may be the best of all, or the best of each species.

    You can also always do (2), again to a subset of all individuals or to a subset (random or sorted) within each species.

    As you correctly anticipate, (4) is also always a possibility, as long as you do have at least two species (it seems things would be a bit broken otherwise).

    Regarding (3) in case you have a species with only one individual? You can't really do it, right?

    There are two things that can help in this situation. First, use a mix of 1 to 4 options. The frequency for each option is normally determined using hyperparameters (as well as the frequency for each type of mutation and so on).

    But here I would actually reconsider your speciation algorithm. Speciation means separating your population into groups, where hopefully more similar individuals are grouped together. There are different ways in which you can do this, and you can re-examine your species with different frequencies as well (you can reset your species every generation!). It does not seem very efficient if your clustering algorithm (because speciation is a type of clustering) is returning species with one or even zero individuals. So this is where I would actually work!

    As a final note, I remember a full NEAT implementation is no basic project. I would recommend not trying to implement this on your own. I think it is a better use of your time to work with a well-established implementation, so you can focus on understanding how things work and how to adapt them for your needs, and not so much on bugs and other implementation details.