{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/deep-q-learning-from-demonstrations","title":"Deep Q-learning from Demonstrations","arxiv_id":"1704.03732","date":"2017-04-12","proceeding":null,"authors":["Todd Hester","Matej Vecerik","Olivier Pietquin","Marc Lanctot","Tom Schaul","Bilal Piot","Dan Horgan","John Quan","Andrew Sendonaris","Gabriel Dulac-Arnold","Ian Osband","John Agapiou","Joel Z. Leibo","Audrunas Gruslys"],"abstract":"Deep reinforcement learning (RL) has achieved several high profile successes\nin difficult decision-making problems. However, these algorithms typically\nrequire a huge amount of data before they reach reasonable performance. In\nfact, their performance during learning can be extremely poor. This may be\nacceptable for a simulator, but it severely limits the applicability of deep RL\nto many real-world tasks, where the agent must learn in the real environment.\nIn this paper we study a setting where the agent may access data from previous\ncontrol of the system. We present an algorithm, Deep Q-learning from\nDemonstrations (DQfD), that leverages small sets of demonstration data to\nmassively accelerate the learning process even from relatively small amounts of\ndemonstration data and is able to automatically assess the necessary ratio of\ndemonstration data while learning thanks to a prioritized replay mechanism.\nDQfD works by combining temporal difference updates with supervised\nclassification of the demonstrator's actions. We show that DQfD has better\ninitial performance than Prioritized Dueling Double Deep Q-Networks (PDD DQN)\nas it starts with better scores on the first million steps on 41 of 42 games\nand on average it takes PDD DQN 83 million steps to catch up to DQfD's\nperformance. DQfD learns to out-perform the best demonstration given in 14 of\n42 games. In addition, DQfD leverages human demonstrations to achieve\nstate-of-the-art results for 11 games. Finally, we show that DQfD performs\nbetter than three related algorithms for incorporating demonstration data into\nDQN.","url_abs":"http://arxiv.org/abs/1704.03732v4","url_pdf":"http://arxiv.org/pdf/1704.03732v4.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/AurelianTactics/dqfd-with-keras","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"unanswered"}},{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/LilTwo/DRL-using-PyTorch","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"unanswered"}},{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/asjad99/Reinforcement-Learning","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/hw9603/DQfD-PyTorch","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"unanswered"}},{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/opendilab/DI-engine","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"deep-q-learning-from-demonstrations","repo_url":"https://github.com/tenvinc/CS4246-Project","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"decision-making","task_name":"Decision Making"},{"task_slug":"deep-reinforcement-learning","task_name":"Deep Reinforcement Learning"},{"task_slug":"imitation-learning","task_name":"Imitation Learning"},{"task_slug":"q-learning","task_name":"Q-Learning"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"},{"task_slug":"reinforcement-learning-1","task_name":"Reinforcement Learning (RL)"}],"methods":[{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dqn","method_name":"DQN"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"q-learning","method_name":"Q-Learning"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1704.03732","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}