Opponent Modeling in Deep Reinforcement Learning

Code for the paper "Opponent Modeling in Deep Reinforcement Learning" published in ICML 2016. The main goal is to learn adpative strategies against different opponents in the deep reinforcement learning framework (Deep Q-Network in particular). Currently it's tested only on Linux with CPU.

Dependencies

• Torch. See installation instructions here.
• Glove word vectors. Can also be downloaded by make dat/glove/glove.840B.300d.txt.

Data

Please email hhe@umiacs.umd.edu for the quiz bowl dataset with human buzzes.

Experiments

Please look at the targets run_qb and run_soccer in the Makefile. To run the quiz bowl experiments, first we need to train a content model (produce the answers) on a separate dataset. See train_content in Makefile. The models will be written to checkpoint_dir and you want to change it to your path.

TODO

• Currently some targets in the Makefile is more like "notes" and the dependencies need to be fixed.
• Test on GPUs.