{"paper":{"title":"Benchmarking Batch Deep Reinforcement Learning Algorithms","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Many batch deep RL algorithms underperform online DQN and the behavioral policy itself when trained on fixed Atari data from one policy.","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Edoardo Conti, Joelle Pineau, Mohammad Ghavamzadeh, Scott Fujimoto","submitted_at":"2019-10-03T20:15:55Z","abstract_excerpt":"Widely-used deep reinforcement learning algorithms have been shown to fail in the batch setting--learning from a fixed data set without interaction with the environment. Following this result, there have been several papers showing reasonable performances under a variety of environments and batch settings. In this paper, we benchmark the performance of recent off-policy and batch reinforcement learning algorithms under unified settings on the Atari domain, with data generated by a single partially-trained behavioral policy. We find that under these conditions, many of these algorithms underper"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"under these conditions, many of these algorithms underperform DQN trained online with the same amount of data, as well as the partially-trained behavioral policy. ... we adapt the Batch-Constrained Q-learning algorithm to a discrete-action setting, and show it outperforms all existing algorithms at this task.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That data generated by a single partially-trained behavioral policy under unified settings produces a representative and fair testbed for comparing batch RL algorithms.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Many batch RL algorithms underperform both online DQN and the behavioral policy on Atari; an adapted discrete-action BCQ outperforms the others tested.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Many batch deep RL algorithms underperform online DQN and the behavioral policy itself when trained on fixed Atari data from one policy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a166490578a6cf086f90d485b3ccaede0b58f6148046db76d51db23d88f1f8ad"},"source":{"id":"1910.01708","kind":"arxiv","version":1},"verdict":{"id":"b0a6fd49-69b5-4df4-b81d-9d426bf70cc4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T19:24:29.450533Z","strongest_claim":"under these conditions, many of these algorithms underperform DQN trained online with the same amount of data, as well as the partially-trained behavioral policy. ... we adapt the Batch-Constrained Q-learning algorithm to a discrete-action setting, and show it outperforms all existing algorithms at this task.","one_line_summary":"Many batch RL algorithms underperform both online DQN and the behavioral policy on Atari; an adapted discrete-action BCQ outperforms the others tested.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That data generated by a single partially-trained behavioral policy under unified settings produces a representative and fair testbed for comparing batch RL algorithms.","pith_extraction_headline":"Many batch deep RL algorithms underperform online DQN and the behavioral policy itself when trained on fixed Atari data from one policy."},"references":{"count":18,"sample":[{"doi":"","year":1907,"title":"Striving for simplicity in off-policy deep reinforcement learning.arXiv preprint arXiv:1907.04543","work_id":"588a4dc4-2821-4960-ab47-4a85390287c3","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Exploration by random network distillation","work_id":"5a87fef6-96e2-4d5b-91ec-1a7c9a43cab9","ref_index":2,"cited_arxiv_id":"1810.12894","is_internal_anchor":true},{"doi":"","year":null,"title":"Dopamine: A research framework for deep reinforcement learning","work_id":"a0c1e29f-e775-49dc-977e-bbbca0daa6d8","ref_index":3,"cited_arxiv_id":"1812.06110","is_internal_anchor":true},{"doi":"","year":null,"title":"Bellemare, and R´ emi Munos","work_id":"f3531420-1352-4be7-b9b7-229fb79f889d","ref_index":4,"cited_arxiv_id":"1710.10044","is_internal_anchor":true},{"doi":"","year":2052,"title":"Off-policy deep reinforcement learning without exploration","work_id":"9bd09ac4-5e91-4d50-9e9b-38436443b511","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":18,"snapshot_sha256":"cf15819122f07fb9170bf0b70ce830d22d3a93022edc2c398111ffc6da057e4c","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}