Home

Awesome

Human-Level Control through Deep Reinforcement Learning

Tensorflow implementation of Human-Level Control through Deep Reinforcement Learning.

model

This implementation contains:

  1. Deep Q-network and Q-learning
  2. Experience replay memory
    • to reduce the correlations between consecutive updates
  3. Network for Q-learning targets are fixed for intervals
    • to reduce the correlations between target and predicted Q-values

Requirements

Usage

First, install prerequisites with:

$ pip install tqdm gym[all]

To train a model for Breakout:

$ python main.py --env_name=Breakout-v0 --is_train=True
$ python main.py --env_name=Breakout-v0 --is_train=True --display=True

To test and record the screen with gym:

$ python main.py --is_train=False
$ python main.py --is_train=False --display=True

Results

Result of training for 24 hours using GTX 980 ti.

best

Simple Results

Details of Breakout with model m2(red) for 30 hours using GTX 980 Ti.

tensorboard

Details of Breakout with model m3(red) for 30 hours using GTX 980 Ti.

tensorboard

Detailed Results

[1] Action-repeat (frame-skip) of 1, 2, and 4 without learning rate decay

A1_A2_A4_0.00025lr

[2] Action-repeat (frame-skip) of 1, 2, and 4 with learning rate decay

A1_A2_A4_0.0025lr

[1] & [2]

A1_A2_A4_0.00025lr_0.0025lr

[3] Action-repeat of 4 for DQN (dark blue) Dueling DQN (dark green) DDQN (brown) Dueling DDQN (turquoise)

The current hyper parameters and gradient clipping are not implemented as it is in the paper.

A4_duel_double

[4] Distributed action-repeat (frame-skip) of 1 without learning rate decay

A1_0.00025lr_distributed

[5] Distributed action-repeat (frame-skip) of 4 without learning rate decay

A4_0.00025lr_distributed

References

License

MIT License.