{"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/q-prop-sample-efficient-policy-gradient-with","title":"Q-Prop: Sample-Efficient Policy Gradient with An Off-Policy Critic","arxiv_id":"1611.02247","date":"2016-11-07","proceeding":null,"authors":["Shixiang Gu","Timothy Lillicrap","Zoubin Ghahramani","Richard E. Turner","Sergey Levine"],"abstract":"Model-free deep reinforcement learning (RL) methods have been successful in a\nwide variety of simulated domains. However, a major obstacle facing deep RL in\nthe real world is their high sample complexity. Batch policy gradient methods\noffer stable learning, but at the cost of high variance, which often requires\nlarge batches. TD-style methods, such as off-policy actor-critic and\nQ-learning, are more sample-efficient but biased, and often require costly\nhyperparameter sweeps to stabilize. In this work, we aim to develop methods\nthat combine the stability of policy gradients with the efficiency of\noff-policy RL. We present Q-Prop, a policy gradient method that uses a Taylor\nexpansion of the off-policy critic as a control variate. Q-Prop is both sample\nefficient and stable, and effectively combines the benefits of on-policy and\noff-policy methods. We analyze the connection between Q-Prop and existing\nmodel-free algorithms, and use control variate theory to derive two variants of\nQ-Prop with conservative and aggressive adaptation. We show that conservative\nQ-Prop provides substantial gains in sample efficiency over trust region policy\noptimization (TRPO) with generalized advantage estimation (GAE), and improves\nstability over deep deterministic policy gradient (DDPG), the state-of-the-art\non-policy and off-policy methods, on OpenAI Gym's MuJoCo continuous control\nenvironments.","url_abs":"http://arxiv.org/abs/1611.02247v3","url_pdf":"http://arxiv.org/pdf/1611.02247v3.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":"q-prop-sample-efficient-policy-gradient-with","repo_url":"https://github.com/shaneshixiang/rllabplusplus","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"NOASSERTION"}},{"paper_slug":"q-prop-sample-efficient-policy-gradient-with","repo_url":"https://github.com/brain-research/mirage-rl","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null}],"tasks":[{"task_slug":"continuous-control","task_name":"Continuous Control"},{"task_slug":"deep-reinforcement-learning","task_name":"Deep Reinforcement Learning"},{"task_slug":"mujoco","task_name":"MuJoCo"},{"task_slug":"policy-gradient-methods","task_name":"Policy Gradient Methods"},{"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)"},{"task_slug":"continuous-control","task_name":"continuous-control"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1611.02247","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}