{"id":"arxiv:2411.03820","title":"Beyond The Rainbow: High Performance Deep Reinforcement Learning on a Desktop PC","date":"2024-11-06","section":"Reinforcement learning","quality":0.211995,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/","source_url":"https://arxiv.org/abs/2411.03820","text_available":true,"authors":["Tyler Clark","Mark Towers","Christine Evers","Jonathon Hare"],"outgoing":[{"source":"arxiv:2411.03820","target":"arxiv:1509.06461","evidence":"title","id":"arxiv:1509.06461","title":"Deep Reinforcement Learning with Double Q-learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a2f115ac879cd2efb9548baf3ecdcb925d2492f03b4b0940720f6e5b0840715b/"},{"source":"arxiv:2411.03820","target":"arxiv:1511.05952","evidence":"id","id":"arxiv:1511.05952","title":"Prioritized Experience Replay","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c88bf0c11c1817dcc457c001dfef5cd120a0946c8a009a0fe40648dae03eb7d4/"},{"source":"arxiv:2411.03820","target":"arxiv:1511.06581","evidence":"title","id":"arxiv:1511.06581","title":"Dueling Network Architectures for Deep Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/0a5f0593cf943c782ba38ae2fc386f953f859b9d3fc9722e8f17f67ca9cf50b5/"},{"source":"arxiv:2411.03820","target":"arxiv:1707.06887","evidence":"title","id":"arxiv:1707.06887","title":"A Distributional Perspective on Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96b040bb771d28ca3254b5519a2191c3e958a35dfdd60075344ac86af6a590e1/"},{"source":"arxiv:2411.03820","target":"arxiv:1710.02298","evidence":"title","id":"arxiv:1710.02298","title":"Rainbow: Combining Improvements in Deep Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4183a5aecdb6558733bd0b2028e6197904c8b46173181bf6c8270fb6cbcfc973/"},{"source":"arxiv:2411.03820","target":"arxiv:2011.14826","evidence":"title","id":"arxiv:2011.14826","title":"Revisiting Rainbow: Promoting more Insightful and Inclusive Deep Reinforcement Learning Research","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2bf20803a5d9e686543fe5f3967e69b5aa481ba29ff91c9ca3fbeb147ab8c0aa/"},{"source":"arxiv:2411.03820","target":"arxiv:2301.04104","evidence":"title","id":"arxiv:2301.04104","title":"Mastering Diverse Domains through World Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/266d59e63d2f705efdc2eb7ad93f5126b7092b4450efd0d8c17099c51574b176/"},{"source":"arxiv:2411.03820","target":"arxiv:2305.19452","evidence":"title","id":"arxiv:2305.19452","title":"Bigger, Better, Faster: Human-level Atari with human-level efficiency","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/727a7ee43fd005401d34d4591a2f49d85175f9366ebf2c9e7a862b340787614e/"}],"incoming":[],"text_parts":[{"start":0,"end":15319,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html"},{"start":15319,"end":31160,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html"},{"start":31160,"end":47074,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html"},{"start":47074,"end":48000,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-4.html"}],"contents":[{"title":"Beyond The Rainbow: High Performance Deep Reinforcement Learning on a Desktop PC","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"Abstract","offset":84,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"Keywords:","offset":1078,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"1 Introduction","offset":1096,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"2 Background","offset":5963,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"2.1 RL Problem Formulation","offset":6107,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"2.2 Deep Q-Learning (DQN)","offset":6906,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"2.3 Rainbow DQN and Improvements to DQN","offset":8147,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"3 Beyond the Rainbow - Extensions and Improvements","offset":11050,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"3.1 Extensions","offset":11856,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-1.html#text"},{"title":"3.2 Hyperparameters","offset":16066,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"4 Evaluation","offset":17718,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"4.1 Atari and Procgen Performance","offset":18078,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"4.2 Applying BTR to Modern Games","offset":19760,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"5 Analysis","offset":21031,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"5.1 Ablations Studies","offset":21327,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"5.2 What are the effects of BTR’s components?","offset":22813,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"6 Related Work","offset":24967,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"7 Conclusions","offset":26967,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Impact Statement","offset":28207,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Acknowledgments","offset":28593,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Appendix A Full Results Tables","offset":29327,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Appendix B Full Results Graphs","offset":29362,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Appendix C Additional Ablations","offset":29397,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Appendix D Hyperparameters","offset":29433,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"D.1 Environment Details","offset":29464,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"D.2 Algorithm Hyperparameters","offset":29493,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"D.3 Clarity of the terms Frames, Steps and Transitions","offset":29528,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"Appendix E Beyond The Rainbow Architecture & Loss Function","offset":30472,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"E.1 Architecture","offset":30535,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-2.html#text"},{"title":"E.2 Loss Function","offset":32291,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"E.3 Analysis Confidence Intervals","offset":33580,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix F BTR with and without Epsilon Greedy","offset":33838,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix G Experiment Compute Resources","offset":34901,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"G.1 Our Compute Resources","offset":34945,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"G.2 BTR with Different Hardware","offset":36165,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix H Other Things We Tried","offset":37551,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix I Altered Atari Environment Settings","offset":40294,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix J BTR for Wii Games","offset":41120,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"J.1 Super Mario Galaxy","offset":41989,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"J.2 Mario Kart Wii","offset":43335,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"J.3 Mortal Kombat","offset":45113,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Appendix K Atari-5 Regression Procedure","offset":46382,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"},{"title":"Instructions for reporting errors","offset":47036,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/part-3.html#text"}],"text_code_points":48000,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bcc96ef608c017d3ddb7ded7dc9e3b760390f7df441214eeee08a2586c981ea/source.md"}