AWhat changed
A model need not paint every future image to estimate whether a situation is promising. The Predictron learns an internal sequence of abstract states, rewards and discounts, combines predictions at several depths and trains them against return targets. Its internal states need to support useful value predictions rather than reproduce the visible environment exactly. Silver and colleagues test the approach on maze-related predictions and simulated pool. The contribution is this learned value-prediction architecture and its training design.
Scientific scope
A model need not paint every future image to estimate whether a situation is promising. The Predictron learns an internal sequence of abstract states, rewards and discounts, combines predictions at several depths and trains them against return targets. Its internal states need to support useful value predictions rather than reproduce the visible environment exactly. Silver and colleagues test the approach on maze-related predictions and simulated pool. The contribution is this learned value-prediction architecture and its training design. The main formulation predicts values in uncontrolled reward processes, rather than supplying a general action-conditional planner. A limited pool experiment ranks candidate initial settings; pocketing-event counts are not automatically percentages of successful episodes. Internal consistency supplements observed supervision and does not prove the model represents true physical dynamics. Learned depth weights do not themselves show that unused computation was skipped. Existing temporal-difference ideas, recurrent networks and later planning systems retain separate credit.
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 main formulation predicts values in uncontrolled reward processes, rather than supplying a general action-conditional planner. A limited pool experiment ranks candidate initial settings; pocketing-event counts are not automatically percentages of successful episodes. Internal consistency supplements observed supervision and does not prove the model represents true physical dynamics. Learned depth weights do not themselves show that unused computation was skipped. Existing temporal-difference ideas, recurrent networks and later planning systems retain separate credit.
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