rotorliu/dqn

Deep Q-learning with Caffe on Space Invaders

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:

• https://github.com/mhauskn/dqn
• https://github.com/muupan/dqn-in-the-caffe

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

• https://sites.google.com/a/deepmind.com/dqn/