3 ms·

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 trai

by deepnet 11y ago

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 ?

hardmaru 11y ago
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.
- deepnet 11y ago
  Many Thanks for the untrained version - I am working on visualising the evolving genome based on your RGBA illustration. I read Sutton and Barto on Q-Learning but only 'got' it when I saw it hand-done with matrix math. http://mnemstudio.org/path-finding-q-learning-tutorial.htm http://mnemstudio.org/path-finding-q-learning-tutorial.htm Sutton & Barto suggest Eligibility Traces seem necessary to make Q learn fast enough. http://webdocs.cs.ualberta.ca/~sutton/book/ebook/node79.html http://webdocs.cs.ualberta.ca/~sutton/book/ebook/node79.html Gomez's thesis suggests that genetic methods have an advantage over backprop for learning sequences of actions. http://www.cs.utexas.edu/users/nn/downloads/papers/gomez.phdtr03.pdf http://www.cs.utexas.edu/users/nn/downloads/papers/gomez.phd... 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.
  - hardmaru 11y ago
    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-

by deepnet 11y ago

hardmaru 11y ago

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.

deepnet 11y ago
Many Thanks for the untrained version - I am working on visualising the evolving genome based on your RGBA illustration. I read Sutton and Barto on Q-Learning but only 'got' it when I saw it hand-done with matrix math. http://mnemstudio.org/path-finding-q-learning-tutorial.htm http://mnemstudio.org/path-finding-q-learning-tutorial.htm Sutton & Barto suggest Eligibility Traces seem necessary to make Q learn fast enough. http://webdocs.cs.ualberta.ca/~sutton/book/ebook/node79.html http://webdocs.cs.ualberta.ca/~sutton/book/ebook/node79.html Gomez's thesis suggests that genetic methods have an advantage over backprop for learning sequences of actions. http://www.cs.utexas.edu/users/nn/downloads/papers/gomez.phdtr03.pdf http://www.cs.utexas.edu/users/nn/downloads/papers/gomez.phd... 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.
- hardmaru 11y ago
  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-