Now, this is very interesting. I read the “Playing Atari with Deep Reinforcement Learning” paper two days ago I decided to do a little research about how this could be applied to trading (sort of a feasibility study). Searching the web for a simpler example I found the “ConvNetJS Deep Q Learning Demo” demo. I though this could serve as a basis for a prototype and then I came across your demo. Hope you will soon post the code on GitHub so that we can fiddle with it and share the findings. A trader never stops researching new ideas. In my opinion, I would not bother incorporating indicators (at least at this point). The OHLC data is more than enough. The efforts needs to be directed to the DQN algorithm and its parameters.
Nice job Peter.