in that they sometimes try to place a stone where one already exists, or they try to place a stone in a place that is considered illegal. when my program encounters such a file, it spits out the error that happened and then continues on with the next file. it's not a big deal. if you want, you can remove that particular game from the directory so it won't happen again. There are enough games so that if you delete one or two it's not a big deal =)... read more
i forgot. Instead of having the source files and the game files separated from one another, i've combined them in this release. i'm gonna also upload a version that doesn't have the games (if you already have them), but you'll need to note where in the directory tree they need to be.
project_dir\Trainer\nn_test_one
and
project_dir\Trainer\nn_test_two
should have 10 randomly chosen games.
project_dir\Trainer\nn_train
should have the rest of the games.
so yeah, i posted on comp.ai.neural-nets today for help on my NN, hopefully some of the visitors are from there. i uploaded the most current version of the source today, although it's not THAT much of an improvement over 0.4, i wanted people to be able to look at the current code so they can help me out. see the release notes for (marginal) details.
but i'm working on it again. one thing that pissed me off is that i just now found out that my way of storing go moves as strings is completely different than the way everybody else does it. i'm not sure if i feel like changing that throughout my whole program at this moment, but i *should* do it at some point.
i've been gone for a week, so not much progress has been made. the Neural Network class still isn't working - i think i'm missing a negative sign somewhere. The error values are getting calculated correctly, but the weight adjustment seems to go in the wrong direction.
I'll have time to start working on the project again sunday night...