{"url":"/method/impala","slug":"impala","name":"IMPALA","full_name":"IMPALA","full_name_withheld":false,"description_markdown":"**IMPALA**, or the **Importance Weighted Actor Learner Architecture**, is an off-policy actor-critic framework that decouples acting from learning and learns from experience trajectories using [V-trace](https://paperswithcode.com/method/v-trace). Unlike the popular [A3C](https://paperswithcode.com/method/a3c)-based agents, in which workers communicate gradients with respect to the parameters of the policy to a central parameter server, IMPALA actors communicate trajectories of experience (sequences of states, actions, and rewards) to a centralized learner. Since the learner in IMPALA has access to full trajectories of experience we use a GPU to perform updates on mini-batches of trajectories while aggressively parallelising all time independent operations. \r\n\r\nThis type of decoupled architecture can achieve very high throughput. However, because the policy used to generate a trajectory can lag behind the policy on the learner by several updates at the time of gradient calculation, learning becomes off-policy. The V-trace off-policy actor-critic algorithm is used to correct for this harmful discrepancy.","description_state":"present","introduced_year":null,"introduced_by":{"title":"IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures","paper":"/paper/impala-scalable-distributed-deep-rl-with","first_author":"Lasse Espeholt","n_authors":12,"url_abs":null,"archive_paper_url":"https://paperswithcode.com/paper/impala-scalable-distributed-deep-rl-with"},"source":{"url":"http://arxiv.org/abs/1802.01561v3","title":"IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures","url_on_a_paper_host":true},"code_snippet_url":null,"code_snippet_url_on_a_code_host":false,"categories":[{"area":"General","area_id":"general","collection":"Distributed Methods","url":"/methods/category/distributed-methods","pwc_aliases":[]},{"area":"Reinforcement Learning","area_id":"reinforcement-learning","collection":"Distributed Reinforcement Learning","url":"/methods/category/distributed-reinforcement-learning","pwc_aliases":[]},{"area":"Reinforcement Learning","area_id":"reinforcement-learning","collection":"Policy Gradient Methods","url":"/methods/category/policy-gradient-methods","pwc_aliases":[]}],"n_papers_tagged":16,"archive_num_papers":16,"papers_newest_first":[{"paper":"/paper/world-model-agents-with-change-based","title":"World Model Agents with Change-Based Intrinsic Motivation","date":"2025-03-26","arxiv_id":"2503.21047","n_code_links":1,"syntology":null},{"paper":"/paper/cleanba-a-reproducible-and-efficient","title":"Cleanba: A Reproducible and Efficient Distributed Reinforcement Learning Platform","date":"2023-09-29","arxiv_id":"2310.00036","n_code_links":1,"syntology":{"ran":4,"of":8,"unverified":4,"pointer_only":8}},{"paper":"/paper/exploring-the-promise-and-limits-of-real-time","title":"Exploring the Promise and Limits of Real-Time Recurrent Learning","date":"2023-05-30","arxiv_id":"2305.19044","n_code_links":1,"syntology":null},{"paper":null,"title":"DoMo-AC: Doubly Multi-step Off-policy Actor-Critic Algorithm","date":"2023-05-29","arxiv_id":"2305.18501","n_code_links":0,"syntology":null},{"paper":"/paper/sharing-lifelong-reinforcement-learning","title":"Sharing Lifelong Reinforcement Learning Knowledge via Modulating Masks","date":"2023-05-18","arxiv_id":"2305.10997","n_code_links":2,"syntology":null},{"paper":"/paper/lifelong-reinforcement-learning-with","title":"Lifelong Reinforcement Learning with Modulating Masks","date":"2022-12-21","arxiv_id":"2212.11110","n_code_links":4,"syntology":null},{"paper":"/paper/envpool-a-highly-parallel-reinforcement","title":"EnvPool: A Highly Parallel Reinforcement Learning Environment Execution Engine","date":"2022-06-21","arxiv_id":"2206.10558","n_code_links":3,"syntology":{"ran":1,"of":5,"unverified":4,"pointer_only":0}},{"paper":null,"title":"Semantic Exploration from Language Abstractions and Pretrained Representations","date":"2022-04-08","arxiv_id":"2204.05080","n_code_links":0,"syntology":null},{"paper":null,"title":"A Distributed Deep Reinforcement Learning Technique for Application Placement in Edge and Fog Computing Environments","date":"2021-10-24","arxiv_id":"2110.12415","n_code_links":0,"syntology":null},{"paper":"/paper/macrpo-multi-agent-cooperative-recurrent","title":"MACRPO: Multi-Agent Cooperative Recurrent Policy Optimization","date":"2021-09-02","arxiv_id":"2109.00882","n_code_links":1,"syntology":null},{"paper":"/paper/particle-based-adaptive-discretization-for","title":"PFPN: Continuous Control of Physically Simulated Characters using Particle Filtering Policy Network","date":"2020-03-16","arxiv_id":"2003.06959","n_code_links":1,"syntology":null},{"paper":null,"title":"A Self-Tuning Actor-Critic Algorithm","date":"2020-02-28","arxiv_id":"2002.12928","n_code_links":0,"syntology":null},{"paper":null,"title":"IMPACT: Importance Weighted Asynchronous Architectures with Clipped Target Networks","date":"2019-11-30","arxiv_id":"1912.00167","n_code_links":0,"syntology":null},{"paper":"/paper/torchbeast-a-pytorch-platform-for-distributed","title":"TorchBeast: A PyTorch Platform for Distributed RL","date":"2019-10-08","arxiv_id":"1910.03552","n_code_links":3,"syntology":{"ran":4,"of":9,"unverified":5,"pointer_only":3}},{"paper":null,"title":"Towards Combining On-Off-Policy Methods for Real-World Applications","date":"2019-04-24","arxiv_id":"1904.10642","n_code_links":0,"syntology":null},{"paper":"/paper/impala-scalable-distributed-deep-rl-with","title":"IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures","date":"2018-02-05","arxiv_id":"1802.01561","n_code_links":24,"syntology":{"ran":16,"of":34,"unverified":18,"pointer_only":3}}],"papers_shown":16,"tasks":[{"task":"/task/reinforcement-learning","name":"Reinforcement Learning","papers":10},{"task":"/task/reinforcement-learning-1","name":"Reinforcement Learning (RL)","papers":8},{"task":"/task/reinforcement-learning-2","name":"reinforcement-learning","papers":8},{"task":"/task/atari-games","name":"Atari Games","papers":2},{"task":"/task/continuous-control","name":"Continuous Control","papers":2},{"task":"/task/deep-reinforcement-learning","name":"Deep Reinforcement Learning","papers":2},{"task":"/task/lifelong-learning","name":"Lifelong learning","papers":2},{"task":"/task/openai-gym","name":"OpenAI Gym","papers":2},{"task":"/task/continuous-control","name":"continuous-control","papers":2},{"task":"/task/diagnostic","name":"Diagnostic","papers":1},{"task":"/task/edge-computing","name":"Edge-computing","papers":1},{"task":"/task/image-captioning","name":"Image Captioning","papers":1},{"task":"/task/mujoco","name":"MuJoCo","papers":1},{"task":null,"name":"Position","papers":1},{"task":"/task/transfer-learning","name":"Transfer Learning","papers":1}],"tasks_shown":15,"n_tasks":15,"usage_by_year":[{"year":"2018","papers":1},{"year":"2019","papers":3},{"year":"2020","papers":2},{"year":"2021","papers":2},{"year":"2022","papers":3},{"year":"2023","papers":4},{"year":"2025","papers":1}],"row_source":"methods_table","archive":{"source":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","archive_url":"https://paperswithcode.com/method/impala"},"syntology_read_at":"2026-09-24T18:15:14+00:00"}