Use a single neural network for each Agent

[xanderdunn]

I broke it up into 9 neural nets for each agent: One for each location on the board. This simplifies the function that any given neural net needs to approximate, but it shouldn't be necessary.