Awesome

An Optimistic Perspective on Offline Reinforcement Learning (ICML, 2020)

This project provides the open source implementation using the Dopamine framework for running experiments mentioned in An Optimistic Perspective on Offline Reinforcement Learning. In this work, we use the logged experiences of a DQN agent for training off-policy agents (shown below) in an offline setting (i.e., batch RL) without any new interaction with the environment during training. Refer to offline-rl.github.io for the project page.

<img src="https://i.imgur.com/Ntgcecq.png" width="95%" alt="Architechture of different off-policy agents" >

Important notes on Atari ROM versions

The DQN replay dataset is generated using a legacy set of Atari ROMs specified in atari-py<=0.2.5, which is different from the ones specified in atari-py>=0.2.6 or in recent versions of ale-py. To avoid train/evaluation mismatches, it is important to use atari-py<=0.2.5 and also gym<=0.19.0, as higher versions of gym no longer support atari-py.

Alternatively, if you prefer to use recent versions of ale-py and gym, you can manually download the legacy ROMs from atari-py==0.2.5 and specify the ROM paths in ale-py. For example, assuming atari_py_rom_breakout is the path to the downloaded ROM file breakout.bin, you can do the following before creating the gym environment:

import ale_py.roms
ale_py.roms.Breakout = atari_py_rom_breakout

Note that this is an ad-hoc trick to circumvent the md5 checks in ale-py<=0.7.5 and it may not work in future versions of ale-py. Do not use this solution unless you know what you are doing.

How to train offline agents on 50M dataset without RAM errors?

Please refer to https://github.com/google-research/batch_rl/issues/10.

JAX codebase

https://github.com/google/dopamine/tree/master/dopamine/labs/offline_rl.

DQN Replay Dataset (Logged DQN data)

The DQN Replay Dataset was collected as follows: We first train a DQN agent, on all 60 Atari 2600 games with sticky actions enabled for 200 million frames (standard protocol) and save all of the experience tuples of (observation, action, reward, next observation) (approximately 50 million) encountered during training.

This logged DQN data can be found in the public GCP bucket gs://atari-replay-datasets which can be downloaded using gsutil. To install gsutil, follow the instructions here.

After installing gsutil, run the command to copy the entire dataset:

gsutil -m cp -R gs://atari-replay-datasets/dqn ./

To run the dataset only for a specific Atari 2600 game (e.g., replace GAME_NAME by Pong to download the logged DQN replay datasets for the game of Pong), run the command:

gsutil -m cp -R gs://atari-replay-datasets/dqn/[GAME_NAME] ./

This data can be generated by running the online agents using batch_rl/baselines/train.py for 200 million frames (standard protocol). Note that the dataset consists of approximately 50 million experience tuples due to frame skipping (i.e., repeating a selected action for k consecutive frames) of 4. The stickiness parameter is set to 0.25, i.e., there is 25% chance at every time step that the environment will execute the agent's previous action again, instead of the agent's new action.

Some Publications using DQN Replay Dataset (please open a pull request for missing entries):

• Revisiting Fundamentals of Experience Replay
• RL Unplugged: A Suite of Benchmarks for Offline Reinforcement Learning
• Conservative Q-Learning for Offline Reinforcement Learning
• Implicit Under-Parameterization Inhibits Data-Efficient Deep Reinforcement Learning
• Acme: A new framework for distributed reinforcement learning
• Regularized Behavior Value Estimation
• Online and Offline Reinforcement Learning by Planning with a Learned Model
• Provable Representation Learning for Imitation with Contrastive Fourier Features
• Decision Transformer: Reinforcement Learning via Sequence Modeling
• DR3: Value-Based Deep Reinforcement Learning Requires Explicit Regularization
• Pretraining Representations for Data-Efficient Reinforcement Learning
• Multi-Game Decision Transformers
• Offline Q-Learning on Diverse Multi-Task Data Both Scales And Generalizes

Asymptotic Performance of offline agents on Atari-replay dataset

Installation

Install the dependencies below, based on your operating system, and then install Dopamine, e.g.

pip install git+https://github.com/google/dopamine.git

Finally, download the source code for batch RL, e.g.

git clone https://github.com/google-research/batch_rl.git

Ubuntu

If you don't have access to a GPU, then replace tensorflow-gpu with tensorflow in the line below (see Tensorflow instructions for details).

sudo apt-get update && sudo apt-get install cmake zlib1g-dev
pip install absl-py atari-py gin-config gym opencv-python tensorflow-gpu

Mac OS X

brew install cmake zlib
pip install absl-py atari-py gin-config gym opencv-python tensorflow

Running Tests

Assuming that you have cloned the batch_rl repository, follow the instructions below to run unit tests.

Basic test

You can test whether basic code is working by running the following:

cd batch_rl
python -um batch_rl.tests.atari_init_test

Test for training an agent with fixed replay buffer

To test an agent using a fixed replay buffer, first generate the data for the Atari 2600 game of Pong to $DATA_DIR.

export DATA_DIR="Insert directory name here"
mkdir -p $DATA_DIR/Pong
gsutil -m cp -R gs://atari-replay-datasets/dqn/Pong/1 $DATA_DIR/Pong

Assuming the replay data is present in $DATA_DIR/Pong/1/replay_logs, run the FixedReplayDQNAgent on Pong using the logged DQN data:

cd batch_rl
python -um batch_rl.tests.fixed_replay_runner_test \
  --replay_dir=$DATA_DIR/Pong/1

Training batch agents on DQN data

The entry point to the standard Atari 2600 experiment is batch_rl/fixed_replay/train.py. Run the batch DQN agent using the following command:

python -um batch_rl.fixed_replay.train \
  --base_dir=/tmp/batch_rl \
  --replay_dir=$DATA_DIR/Pong/1 \
  --gin_files='batch_rl/fixed_replay/configs/dqn.gin'

By default, this will kick off an experiment lasting 200 training iterations (equivalent to experiencing 200 million frames for an online agent).

To get finer-grained information about the process, you can adjust the experiment parameters in batch_rl/fixed_replay/configs/dqn.gin, in particular by increasing the FixedReplayRunner.num_iterations to see the asymptotic performance of the batch agents. For example, run the batch REM agent for 800 training iterations on the game of Pong using the following command:

python -um batch_rl.fixed_replay.train \
  --base_dir=/tmp/batch_rl \
  --replay_dir=$DATA_DIR/Pong/1 \
  --agent_name=multi_head_dqn \
  --gin_files='batch_rl/fixed_replay/configs/rem.gin' \
  --gin_bindings='FixedReplayRunner.num_iterations=1000' \
  --gin_bindings='atari_lib.create_atari_environment.game_name = "Pong"'

More generally, since this code is based on Dopamine, it can be easily configured using the gin configuration framework.

Dependencies

The code was tested under Ubuntu 16 and uses these packages:

• tensorflow-gpu>=1.13
• absl-py
• atari-py
• gin-config
• opencv-python
• gym
• numpy

The python version upto 3.7.9 has been reported to work.

Citing

If you find this open source release useful, please reference in your paper:

Agarwal, R., Schuurmans, D. & Norouzi, M.. (2020). An Optimistic Perspective on Offline Reinforcement Learning International Conference on Machine Learning (ICML).

@inproceedings{agarwal2020optimistic,
  title={An Optimistic Perspective on Offline Reinforcement Learning},
  author={Agarwal, Rishabh and Schuurmans, Dale and Norouzi, Mohammad},
  journal={International Conference on Machine Learning},
  year={2020}
}

Note: A previous version of this work was titled "Striving for Simplicity in Off Policy Deep Reinforcement Learning" and was presented as a contributed talk at NeurIPS 2019 DRL Workshop.

Disclaimer: This is not an official Google product.