For example, I assume that a dogs past performance must be some indicator of its chances in the next race, but how do I account for the chances of dogs who didn't complete their last race? What weighting is the last race worth, compared to the ones before (perhaps it had a bad race, but on the whole is running well).
I just don't know how to optimise for those sort of things. I have a rough idea that some combination of genetic programming and GA could help - it would be an interesting challenge to builds software that knew how to apply a selection of mathematical functions to my data, and then breed the results like a GA. But it's tricky, I'd have thought.
I've been treating the ML classifiers and learners as something of a black box, perhaps a more rigorous approach is required.
yeah, I'd envisaged using the accuracy of the neural net as the fitness function for a GA that mutates input parameters. It's another layer of complexity, and I've no clue how you'd start, but it seems like it would work.
In other words - use a GA to select features, using how well the NN trained on that set of features performs as the fitness function.