Which is just subtracting zero from the result. Not sure why. The only reason I know for doing that is to cast a string to a number without the function overhead of parseInt(). Furthermore, I don't actually see any code using that function.
lol - apologies about my messy code guys. It's just hacked together as a rough sketch to see if it'll work. as you can see my web coding skillz still needs work. I put it up on github though (be warned about messy hackish code):
I found the code clear and comprehensible - your genetic library for convnet.js looks useful - have you seen Karpathy's reinforce.js ?
I notice you left a training flag - I would like to watch the nets train so definitely going to have a go.
And thanks for introducing me to CNE, I am reading John Gomez's thesis right now - very very interesting, looks to have a big advantage over Q-learning.
In your blog post you compare the 'DNA' of multiple training episodes, the similar weights kind of suggests that you have exhaustively searched the problem space ?
Thanks - thinking of putting up a version that has not been pretrained, so users can see the training over several generations. I think I have one lying around - remind me on twitter if I forget.
I'm also trying to learn 'deep q learning' (q-learning with a neural net as the q function) and other cool stuff that has recently been developed.
btw, I think your previous post with the 'brain' of 140 numbers is messing up this thread a bit on my browser for some reason and the comments not wrap, wonder if there's a way to fix it.
Deep Q Learning is brilliant - evolving the Q function would certainly be very interesting - currently trying to get the ALE sending frames to my convnet :)
mods have helped fix the errant brain code - thanks for the heads up.
cool! Let's talk further offline later- I would be interested to know what you end up doing with it, and pick your (biological) brain for ideas. Ping me on Twitter message later-
Hope it didn't seem like I was trying to criticize. The app is really cool, and thanks for open-sourcing it. I'd be really interested in seeing the version that doesn't start pre-trained.
