https://www.youtube.com/watch?v=fevMOp5TDQs
http://www.denizyuret.com/2015/03/alec-radfords-animations-for.html
artari is not a MDP, but MDP method works well. or use RNN
in many domains, people end up using RNN to represent q-function.
replay really makes a difference!!!
should the two network have different set of hyperparameter? just like a group of workers with different kinds of personality? will the collaboration help?