rldb 0.0.0

Performances of Reinforcement Learning Agents

rldb

Database of RL algorithms

Atari Space Invaders Scores	MuJoCo Walker2d Scores

Examples

You can use rldb.find_all({}) to retrieve all existing entries in rldb.

import rldb


all_entries = rldb.find_all({})

You can also filter entries by specifying key-value pairs that the entry must match:

import rldb


dqn_entries = rldb.find_all({'algo-nickname': 'DQN'})
breakout_noop_entries = rldb.find_all({
    'env-title': 'atari-breakout',
    'env-variant': 'No-op start',
})

You can also use rldbl.find_one(filter_dict) to find one entry that matches the key-value pair specified in filter_dict:

import rldb
import pprint


entry = rldb.find_one({
    'env-title': 'atari-pong',
    'algo-title': 'Human',
})
pprint.pprint(entry)

Output

{
    'algo-nickname': 'Human',
    'algo-title': 'Human',
    'env-title': 'atari-pong',
    'env-variant': 'No-op start',
    'score': 14.6,
    'source-arxiv-id': '1511.06581',
    'source-arxiv-version': 3,
    'source-authors': [   'Ziyu Wang',
                          'Tom Schaul',
                          'Matteo Hessel',
                          'Hado van Hasselt',
                          'Marc Lanctot',
                          'Nando de Freitas'],
    'source-bibtex': '@article{DBLP:journals/corr/WangFL15,\n'
                     '    author    = {Ziyu Wang and\n'
                     '                 Nando de Freitas and\n'
                     '                 Marc Lanctot},\n'
                     '    title     = {Dueling Network Architectures for Deep '
                     'Reinforcement Learning},\n'
                     '    journal   = {CoRR},\n'
                     '    volume    = {abs/1511.06581},\n'
                     '    year      = {2015},\n'
                     '    url       = {http://arxiv.org/abs/1511.06581},\n'
                     '    archivePrefix = {arXiv},\n'
                     '    eprint    = {1511.06581},\n'
                     '    timestamp = {Mon, 13 Aug 2018 16:48:17 +0200},\n'
                     '    biburl    = '
                     '{https://dblp.org/rec/bib/journals/corr/WangFL15},\n'
                     '    bibsource = {dblp computer science bibliography, '
                     'https://dblp.org}\n'
                     '}',
    'source-nickname': 'DuDQN',
    'source-title': 'Dueling Network Architectures for Deep Reinforcement '
                    'Learning'
}

Entry Structure

Here is the format of every entry:

{
    # BASICS
    "source-title": "",
    "source-nickname": "",
    "source-authors": [],

    # MISC.
    "source-bibtex": "",

    # ALGORITHM
    "algo-title": "",
    "algo-nickname": "",
    "algo-source-title": "",

    # SCORE
    "env-title": "",
    "score": 0,
}

• source-title is the full title of the source of the score: it can be the title of the paper or GitHub repository title. source-nickname is a popular nickname or acronym for that title if it exists, otherwise it is the same as source-title.
• source-authors are a list of authors or contributors.
• source-bibtex is a BibTeX-format citation.
• algo-title is the full title of the algorithm used. algo-nickname is the nickname or acronym for that algorithm if it exists, otherwise it is the same as algo-nickname.
• algo-source-title is the title of the source of the algorithm. It can and often is different from source-title.

For example, the Space Invaders score of Asynchronous Advantage Actor Critic (A3C) algorithm in the Noisy Networks for Exploration (NoisyNet) paper is represented by the following entry:

{
    #  BASICS
    "source-title": "Noisy Networks for Exploration",
    "source-nickname": "NoisyNet",
    "source-authors": [
        "Meire Fortunato",
        "Mohammad Gheshlaghi Azar",
        "Bilal Piot",
        "Jacob Menick",
        "Ian Osband",
        "Alex Graves",
        "Vlad Mnih",
        "Remi Munos",
        "Demis Hassabis",
        "Olivier Pietquin",
        "Charles Blundell",
        "Shane Legg",
    ],

    #  ARXIV
    "source-arxiv-id": "1706.10295",
    "source-arxiv-version": 2,

    #  MISC.
    "source-bibtex": """
@article{DBLP:journals/corr/FortunatoAPMOGM17,
    author    = {Meire Fortunato and
                 Mohammad Gheshlaghi Azar and
                 Bilal Piot and
                 Jacob Menick and
                 Ian Osband and
                 Alex Graves and
                 Vlad Mnih and
                 R{\'{e}}mi Munos and
                 Demis Hassabis and
                 Olivier Pietquin and
                 Charles Blundell and
                 Shane Legg},
    title     = {Noisy Networks for Exploration},
    journal   = {CoRR},
    volume    = {abs/1706.10295},
    year      = {2017},
    url       = {http://arxiv.org/abs/1706.10295},
    archivePrefix = {arXiv},
    eprint    = {1706.10295},
    timestamp = {Mon, 13 Aug 2018 16:46:11 +0200},
    biburl    = {https://dblp.org/rec/bib/journals/corr/FortunatoAPMOGM17},
    bibsource = {dblp computer science bibliography, https://dblp.org}
}""",

    # ALGORITHM
    "algo-title": "Asynchronous Advantage Actor Critic",
    "algo-nickname": "A3C",
    "algo-source-title": "Asynchronous Methods for Deep Reinforcement Learning",

    # HYPERPARAMETERS
    "algo-frames": 320 * 1000 * 1000,  # Number of frames

    # SCORE
    "env-title": "atari-space-invaders",
    "env-variant": "No-op start",
    "score": 1034,
    "stddev": 49,
}

Note that, as shown here, the entry can contain additional information.

Sources

Papers

Deep Q-Networks

• Playing Atari with Deep Reinforcement Learning (Mnih et al., 2013)
• Human-level control through deep reinforcement learning (Mnih et al., 2015)
• Deep Recurrent Q-Learning for Partially Observable MDPs (Hausknecht and Stone, 2015)
• Massively Parallel Methods for Deep Reinforcement Learning (Nair et al., 2015)
• Deep Reinforcement Learning with Double Q-learning (Hasselt et al., 2015)
• Prioritized Experience Replay (Schaul et al., 2015)
• Dueling Network Architectures for Deep Reinforcement Learning (Wang et al., 2015)
• Noisy Networks for Exploration (Fortunato et al., 2017)
• A Distributional Perspective on Reinforcement Learning (Bellemare et al., 2017)
• Rainbow: Combining Improvements in Deep Reinforcement Learning (Hessel et al., 2017)
• Distributional Reinforcement Learning with Quantile Regression (Dabney et al., 2017)
• Implicit Quantile Networks for Distributional Reinforcement Learning (Dabney et al., 2018)

Policy Gradients

• Asynchronous Methods for Deep Reinforcement Learning (Mnih et al., 2016)
• Trust Region Policy Optimization (Schulman et al., 2015)
• Proximal Policy Optimization Algorithms (Schulman et al., 2017)
• Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation (Wu et al., 2017)
• Addressing Function Approximation Error in Actor-Critic Methods (Fujimoto et al., 2018)
• IMPALA: Importance Weighted Actor-Learner Architectures (Espeholt et al., 2018)
• The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning (Gruslys et al., 2017)

Exploration

• Exploration by Random Network Distillation (Burda et al., 2018)

Misc.

• Trust-PCL: An Off-Policy Trust Region Method for Continuous Control (Nachum et al., 2017)

Repositories

• OpenAI Baselines
• RL Baselines Zoo