While some features in NEAT are currently missing, others are intentionally excluded or changed.
The original NEAT paper seems to suggest that new genomes produced through crossover should randomly inherit the disablement of a connection gene if it is disabled in either parent, regardless of the relative fitness of the parent carrying the disabled gene.
This leads to a choice between, at one extreme, allowing rather frequent inheritance of disablement from weak parents, or rarely inheriting disablement from fit parents, at the other. Since relative fitness determines which parent passes on excess and disjoint genes, neither option is appealing as both options frequently disrupt effective structure. Attempting to compromise with a fifty percent chance of inheriting disablement from either parent leads to frequently disrupted structure in both of these problematic scenarios. This is caused by an effective double bind due to there being only one probability covering multiple scenarios.
swirlnet gives you flexible options through seven probability settings.
-
Firstly, you have the option of of using these settings to disable the special inheritance rules for gene disablement.
- This causes disablement and enablement to be inherited along with connection weight according to the gene inheritance rules.
-
You can set gene disablement status to always be inherited from the gene of the fitter parent, along with other structural differences.
- In the case of the XOR challenge, this setting and settings with similar probabilities are the most effective options I've tried since they reduce attraction to the the local minimum of creating genomes with every connection gene disabled. Without this type of inheritance, connections often become disabled when a weak genome with an added (and reconnected) node and disabled (short) connection reproduces with a fitter genome. The additional connected node is not inherited from the weaker genome since it is either disjoint or excess; if the disabled connection is randomly inherited from the weaker genome then the result is a new genome with a missing connection. Once all connections are deleted, genomes always create the same output that is situated half way between the correct and incorrect answer for every XOR test case. In early stages of evolution, this disconnected genome may outperform connected genomes. Using the above approach solves this particular issue.
-
By changing four settings in swirlnet you may differentiate the probabilities of inheriting disablement, inheriting enablement or non-interference, each according to whether the fitter or weaker parent has the disabled gene.
-
You may set the gene enablement rate for cases in which both parents' genes are disabled.
-
You can also set the frequency of asexual gene disablement toggling with two settings, one for the rate of enablement, the other for the rate of disablement.
-
Tweaking these settings allows you to follow the same, or a similar, approach as the one outlined in the original NEAT paper.
The additional rules for sexually inherited disablement occur after each connection weight is randomly inherited from either parent.
If two genomes with equal fitness scores are undergoing crossover, one of them is treated as the fitter genome by swirlnet. Excess and disjoint genes are inherited from the 'fitter' parent instead of randomly inheriting disjoint and excess genes from both.
Removing this feature decreased code complexity and increased the rate at which solutions to a prior (albeit naive) version of the XOR problem were found. The prior behavior may be re-added later if found to be useful.
swirlnet offers expanded options for controlling weight mutation compared to suggestions in the original NEAT paper. One has the option of setting the uniform gene perturbation, unique gene perturbation and unique gene weight reset rates. The sum of these rates may be less than 1 in order to allow some weights in a genome undergoing weight mutation to remain unmutated.
swirlnet uses Gaussian distributions to set new weights and perturb existing ones.
There are likely many other undocumented ways in which this library differs from other implementations of NEAT.