Deep Q-learning with Caffe on Space Invaders
C++ Python Cuda Protocol Buffer Makefile Matlab Shell
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
docs
examples
include/caffe
matlab/caffe
python
scripts
src
tools
.gitignore
.travis.yml
CONTRIBUTORS.md
INSTALL.md
LICENSE
Makefile
Makefile.config
Makefile.config.dev
Makefile.config.example
README.md
caffe.cloc

README.md

Summary

This was the first open source version of DeepMind's DQN paper. In addition, a crowd-based reward singal was collected which you can use to train your model, available here:

http://aiworld.io/data/space-invaders.html

Details

All reinforcement learning done in Python. In addition, solver.cpp was modified to support online observation of training data with Solver<Dtype>::Solve split into OnlineUpdateSetup, OnlineUpdate, and OnlineForward to set the input of the memory data layer, determine the q-loss in examples/dqn, then optionally backprop depending on whether we are training or just acting.

To use the crowd-reward data, download from above and set the following in your environment:

export INTEGRATE_HUMAN_FEEDBACK=True

Similar projects:

Official improved DQN updated and released Feb 25th built on Torch