Re:Introduction to Monte Carlo Tree Search
by versteegen on Sunday September 13, 2015 @10:24AM
(#10211491)
(That was before Google acquired DeepMind, and probably a large part of the reason.)
This uses a deep convolutional NN as an approximation function of the expected value. Interestingly there are only 3 hidden layers in this CNN, and it only gets fed the last 4 frames as state. Of course normally you would feed your approximator a direct encoding of the state. If you wanted to use a NN to evaluate nodes in gametree search you would have to invent some other way to train it. Would love to know whether there is research on that.
https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf