Those with an interest in Genetic Algorithm training...

[Dave Schwartz]

If any of you have an interest in Genetic Algorithm training, you may enjoy following this thread on our BBS as several guys doing training participate in an on-going discussion.

The thread is here:
http://www.horsestreet.com/ubb/Forum4/HTML/002124.html

The current point of the discussion is about in-breeding.

[Handiman]

Sure that isn't about folk from the Ozarks who have moved to Reno and are into Horses? :lol:

Sorry Dave.... Couldn't help myself.

Handi:)

[Dave Schwartz]

Them hill folk ain't got no idee bout no generic algo...nomics.

[LemonDropKid]

Dave,

I read your thread and have used ANN's with GA's to predict horse races. Continuous training would always lead to overfitting. How is this being addressed with your ant hill?

"The Kid"

[Dave Schwartz]

Over fitting is caused by an unrepresentative sample. If the sample is large enough there is simply no way that the "hill" can train to convergence.

IOW, I am just not concerned.

Also, "continuous" is not in the plan either. I was surprised at the training speed of these ants. The train slower than any I have ever built before. But while they learn slower, I believe they learn "deeper."


This is the first time I have ever trained with the goal being to simply drive winners to the top. I am pleased with the output thus far, although I really expected it to be a little more absolute in its selections.

That is, I expected it to only make A's out of 4/5 shots and that they would win at a huge rate. Instead I am finding that the selections are not near as obvious. I handicapped a race from about a month ago that actually put a $60 horse on top as a B and I have seen many races with $15-$30 horses as A's. I don't mean like one a day but certainly a couple each week. And the ROI on such horses (anecdotally) has to be huge.


Dave