{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "title": "UNREAL: auxiliary learning tasks that improve a control representation", "summary": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods.", "boundary": "The base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.", "area": "Reinforcement-learning representation learning", "topic": "Artificial intelligence", "year": 2016, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "humans": [ "Max Jaderberg", "Volodymyr Mnih", "Wojciech Marian Czarnecki", "Tom Schaul", "Joel Z Leibo", "David Silver", "Koray Kavukcuoglu" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [ "deep-q-network" ], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-1611-05397v1", "title": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1", "date": "2016-11-16" } ], "geography": { "country_codes": [ "GB" ], "location_basis": "Original DeepMind byline explicitly London, UK. Work affiliation, not nationality or ownership.", "source_url": "https://arxiv.org/abs/1611.05397v1", "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "basis": "RESEARCH_LOCATION_OR_PAPER_AFFILIATION", "reviewed_on": "2026-10-23", "nationality_inferred": false, "sources": [ { "title": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } ], "coverage_status": "COMPLETE", "country_labels": { "GB": "United Kingdom" } }, "points": "90000", "published_grade": { "id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "entry_id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "what_changed": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods.", "scope": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods. The base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Max Jaderberg", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Volodymyr Mnih", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Wojciech Marian Czarnecki", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Tom Schaul", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Joel Z Leibo", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "David Silver", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Koray Kavukcuoglu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.", "domain": "Reinforcement-learning representation learning" }, "sources": [ { "citation": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1" } ], "statement": "# UNREAL: auxiliary learning tasks that improve a control representation\n\nA game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods.\n\n## Scientific scope\n\nThe base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nMax Jaderberg; Volodymyr Mnih; Wojciech Marian Czarnecki; Tom Schaul; Joel Z Leibo; David Silver; Koray Kavukcuoglu\n\nSource-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.\n\n\n\n## Original sources\n\n- [Reinforcement Learning with Unsupervised Auxiliary Tasks](https://arxiv.org/abs/1611.05397v1)\n", "summary": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods.", "title": "UNREAL: auxiliary learning tasks that improve a control representation" }, "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "source_ref": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:1413647140f874686659c02986e06752676049d0f53bfd1dce4a32e53a562125" }, "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "entry_id": "jaderberg-mnih-czarnecki-et-al-unreal-auxiliary-control-2016", "revision": 1, "title": "UNREAL: auxiliary learning tasks that improve a control representation", "context": "Reinforcement-learning representation learning", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods.", "scope": "A game-playing agent learns more than a final score when it also practices useful auxiliary tasks. UNREAL adds pixel-change control, reward prediction and replayed value learning to an A3C agent, sharing parts of the visual and recurrent representation. These extra learning signals help the agent use its experience more effectively in the reported Labyrinth and Atari experiments. The contribution is the combined auxiliary-learning architecture and its finite evaluation, building on existing actor–critic, replay and off-policy methods. The base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The base task still uses external rewards; “unsupervised” auxiliary learning is not a reward-free general agent. Reported Labyrinth mean performance rises from 54% to 87% on a human-normalized scale, about 1.61 times rather than more than twice. Step-efficiency estimates depend on a specified A3C target, selected runs and aggregation across tasks; they are not general wall-clock savings. Game scores do not measure general intelligence. Some alternative auxiliary losses hurt final performance, and feature-control results are preliminary.", "references": [ { "title": "Reinforcement Learning with Unsupervised Auxiliary Tasks", "url": "https://arxiv.org/abs/1611.05397v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:36265b43-c12a-4f5f-a923-1eb6e58a222c", "version": 1, "content_hash": "sha256:4cf8b67a5d49eba3764b069a422d649152735b7d05ebf608c5540b5af1329da6" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "90000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": null, "reading_status": "SOURCE_LINKED_NOTE" }