teaching mario to play mario

This is a very simple example using reinforceio/tensorforce, tensorflow and openai/gym. This was a small weekend project when DQfD was first published and I thought others might use it as a positive-slope for their experiments getting started. I don't have the access to computing power to let this little guy run for long periods of time. Most of the agents I've trained here make use of my own demonstrations playing mario and they are sparse. I must say, I think the agent is just about as good as I am (that's definitely not saying much). I eventually want to combine this technique with evolution strategies, since both frameworks are very straight-forward, scalable, and easy to implement. It would be interesting to combine the weighted ranking algorithm of evolution strategies with replay prioritization to make the next generation of parameters a function of the demonstrations of past generations (if that makes sense, more on that later).

cnn

cnn+lstm

If you want specific access to more documentation and details, please send me an email and I may be able to help!

NOTE: that this will probably not run on your system because it required a few changes to openai/gym and ppaquette/gym-super-mario, as well as reinforceio/tensorforce :(. I basically steem-rolled the errors that came up (openai/gym changes were due to a depreciation of an Env Wrapper that ppaquette/gym-super-mario used. I chose to remedy this issue by cherrypicking the changes and placing it into ppaquette/gym-super-mario, since it was removed without depreciation warning.)

The reinforceio/tensorforce changes are a little more complicated, since they involved a small DQFDAgent change and a few code fixes - I will most likely submit a pull request for most of these libraries so everyone can use this.

attempt to install

# install/upgrade tensorforce
pip install --upgrade tensorforce
# install gym extensions
pip install gym_pull
# install mario environment
pip install ppaquette_gym_super_mario
python -c "import gym_pull; gym_pull.pull('github.com/ppaquette/gym-super-mario')"

attempt to teach/train and validate

# if your starting from scratch
mkdir agents demos monitors
# now run this, change if necessary
python mario.py ppaquette/meta-SuperMarioBros-v0 \
  -a DQFDAgent -c mario_agent.json -pt \
  -ld ./demos/ -s ./agents/ -m ./monitors/ -mv 100 -D

recording how bad you are at playing the game

python demo.py ppaquette/meta-SuperMarioBros-v0 -s .
# start playing... it loops over when you die and starts a new demo file.