Simple DQN model, seen around 10M frames (iterations)

|

python dqn_atari.py \
  --env Enduro-v0 \
  --gpu 0 \
  --model convnet \
  --train_policy epgreedy \
  --std_img \
  --optimizer adam \
  --learning_rate 0.0001

Simply replace --model convnet with --model dueling_convnet in the above command. Also try out other network architectures in deeprl/networks.py.

Following curves compare the dueling (yellow), double (green) and simple (blue) deep Q networks.

Episode length

Total Reward over 20 episodes

Loss

This work was done as a course assignment for the CMU Deep RL course, so thanks to the instructors for guidance and providing starter code. Also thanks to Achal for hyperparameter suggestions.

[1] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013.

[2] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.

[3] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. 2016.

[4] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado van Hasselt, Marc Lanctot, and Nando de Freitas. Dueling network architectures for deep reinforcement learning. arXiv preprint arXiv:1511.06581, 2015.