Skip to content

Deep Q-Learning (DQN)

Overview

As an extension of the Q-learning, DQN's main technical contribution is the use of replay buffer and target network, both of which would help improve the stability of the algorithm.

Original papers:

Implemented Variants

Variants Implemented Description
dqn.py, docs For classic control tasks like CartPole-v1.
dqn_atari.py, docs For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques.

Below are our single-file implementations of DQN:

dqn.py

The dqn.py has the following features:

  • Works with the Box observation space of low-level features
  • Works with the Discrete action space
  • Works with envs like CartPole-v1

Implementation details

dqn.py includes the 11 core implementation details:

dqn_atari.py

The dqn_atari.py has the following features:

  • For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
  • Works with the Atari's pixel Box observation space of shape (210, 160, 3)
  • Works with the Discrete action space

Usage

poetry install -E atari
python cleanrl/dqn_atari.py --env-id BreakoutNoFrameskip-v4
python cleanrl/dqn_atari.py --env-id PongNoFrameskip-v4

Implementation details

dqn_atari.py is based on (Mnih et al., 2015)1 but presents a few implementation differences:

  1. dqn_atari.py use slightly different hyperparameters. Specifically,
    • dqn_atari.py uses the more popular Adam Optimizer with the --learning-rate=1e-4 as follows: 
      optim.Adam(q_network.parameters(), lr=1e-4)
      whereas (Mnih et al., 2015)1 (Exntended Data Table 1) uses the RMSProp optimizer with --learning-rate=2.5e-4, gradient momentum 0.95, squared gradient momentum 0.95, and min squared gradient 0.01 as follows: 
      optim.RMSprop(
    q_network.parameters(),
    lr=2.5e-4,
    momentum=0.95,
    # ... PyTorch's RMSprop does not directly support
    # squared gradient momentum and min squared gradient
    # so we are not sure what to put here.
)
    • dqn_atari.py uses --learning-starts=80000 whereas (Mnih et al., 2015)1 (Exntended Data Table 1) uses --learning-starts=50000.
    • dqn_atari.py uses --total-timesteps=10000000 (i.e., 10M timesteps = 40M frames because of frame-skipping) whereas (Mnih et al., 2015)1 uses --total-timesteps=12500000 (i.e., 12.5M timesteps = 50M frames) (See "Training details" under "METHODS" on page 6).
    • dqn_atari.py uses --end-e=0.01 (the final exploration epsilon) whereas (Mnih et al., 2015)1 (Exntended Data Table 1) uses --end-e=0.1.
  2. dqn_atari.py use a self-contained evaluation scheme: dqn_atari.py simply reports the episodic returns obtained throughout training, whereas (Mnih et al., 2015)1 is trained with --end-e=0.1 but reported episodic returns using a separate evaluation process with --end-e=0.01 (See "Evaluation procedure" under "METHODS" on page 6).

  1. Mnih, V., Kavukcuoglu, K., Silver, D. et al. Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015). https://doi.org/10.1038/nature14236 ↩↩↩↩↩↩

Back to top