Logarithmic Reinforcement Learning
Перейти к файлу
Mehdi Fatemi 915fda24f4
Merge pull request #10 from microsoft/dependabot/pip/log_dqn_experiments/tensorflow-2.9.3
Bump tensorflow from 2.7.2 to 2.9.3 in /log_dqn_experiments
2022-11-22 12:22:31 -05:00
linear_experiments update legal notes 2019-10-29 12:11:02 -04:00
log_dqn_experiments Bump tensorflow from 2.7.2 to 2.9.3 in /log_dqn_experiments 2022-11-21 23:41:10 +00:00
CODE_OF_CONDUCT.md Initial CODE_OF_CONDUCT.md commit 2019-10-15 10:20:13 -07:00
CONTRIBUTING.md update legal notes 2019-10-29 12:11:02 -04:00
LICENSE update legal notes 2019-10-29 12:11:02 -04:00
NOTICE.md update legal notes 2019-10-29 12:11:02 -04:00
README.md update legal notes 2019-10-29 12:11:02 -04:00
SECURITY.md Initial SECURITY.md commit 2019-10-15 10:20:15 -07:00

README.md

Logarithmic Reinforcement Learning

This repository hosts sample code for the NeurIPS 2019 paper: van Seijen, Fatemi, Tavakoli (2019).

We provide code for the linear experiments of the paper as well as the deep RL Atari 2600 examples (LogDQN).

LICENSE

Microsoft Open Source Code of Conduct

The code for LogDQN has been developed by Arash Tavakoli and the code for the linear experiments has been developed by Harm van Seijen.

Citing

If you use this research in your work, please cite the accompanying paper:

@inproceedings{vanseijen2019logrl,
  title={Using a Logarithmic Mapping to Enable Lower Discount Factors in Reinforcement Learning},
  author={van Seijen, Harm and
          Fatemi, Mehdi and 
          Tavakoli, Arash},
  booktitle={Advances in Neural Information Processing Systems},
  year={2019}
}

Linear Experiments

First navigate to linear_experiments folder.

To create result-files:

python main

To visualize result-files:

python show_results

With the default settings (i.e., keeping main.py unchanged), a scan over different gamma values is performed for a tile-width of 2 for a version of Q-learning without a logarithmic mapping.

All experimental settings can be found at the top of the main.py file. To run the logarithmic-mapping version of Q-learning, set:

agent_settings['log_mapping'] = True

Results of the full scans are provided. To visualize these results for regular Q-learning or logarithmic Q-learning, set filename in show_results.py to full_scan_reg or full_scan_log, respectively.


Logarithmic Deep Q-Network (LogDQN)

This part presents an implementation of LogDQN from van Seijen, Fatemi, Tavakoli (2019).

Instructions

Our implementation of LogDQN builds on Dopamine (Castro et al., 2018), a Tensorflow-based research framework for fast prototyping of reinforcement learning algorithms.

Follow the instructions below to install the LogDQN package along with a compatible version of Dopamine and their dependencies inside a conda environment.

First install Anaconda, and then proceed below.

conda create --name log-env python=3.6 
conda activate log-env

Ubuntu

sudo apt-get update && sudo apt-get install cmake zlib1g-dev
pip install absl-py atari-py gin-config gym opencv-python tensorflow==1.15rc3
pip install git+git://github.com/google/dopamine.git@a59d5d6c68b1a6e790d5808c550ae0f51d3e85ce

Finally, navigate to log_dqn_experiments and install the LogDQN package from source.

cd log_dqn_experiments
pip install .

Training an agent

To run a LogDQN agent,

python -um log_dqn.train_atari \
    --agent_name=log_dqn \
    --base_dir=/tmp/log_dqn \
    --gin_files='log_dqn/log_dqn.gin' \
    --gin_bindings="Runner.game_name = \"Asterix\"" \
    --gin_bindings="LogDQNAgent.tf_device=\"/gpu:0\""

You can set LogDQNAgent.tf_device to /cpu:* for a non-GPU version.