AWhat 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.
Scientific 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
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.
Contributors
- The named authors in the research record · SCIENCE: 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.
Provisional WRS assessment, edition 2026-10-23.6. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction