AWhat changed
An observation-based planner tries to predict what it will see after an action. A Value Prediction Network instead learns action-option-dependent changes in an abstract state and predicts rewards, discounts and future values. It can expand several possible options and back up their estimated returns without generating future images. Oh, Singh and Lee combine this lookahead with multistep learning and evaluate it in a collection task and selected Atari games. This adds controlled planning to the abstract-value approach, rather than re-awarding the earlier uncontrolled Predictron.
Scientific scope
An observation-based planner tries to predict what it will see after an action. A Value Prediction Network instead learns action-option-dependent changes in an abstract state and predicts rewards, discounts and future values. It can expand several possible options and back up their estimated returns without generating future images. Oh, Singh and Lee combine this lookahead with multistep learning and evaluate it in a collection task and selected Atari games. This adds controlled planning to the abstract-value approach, rather than re-awarding the earlier uncontrolled Predictron. The search uses a restricted branching heuristic and hand-designed options. An abstract state is not proved to be an exact representation of every relevant future. Deeper planning beyond the training range is not guaranteed to help. The Atari comparison improves seven of nine games and uses selected top agents, with a frame-skip setting that affects comparisons. These are finite results, not an optimal-policy theorem, universal noise robustness or the later MuZero system.
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 search uses a restricted branching heuristic and hand-designed options. An abstract state is not proved to be an exact representation of every relevant future. Deeper planning beyond the training range is not guaranteed to help. The Atari comparison improves seven of nine games and uses selected top agents, with a frame-skip setting that affects comparisons. These are finite results, not an optimal-policy theorem, universal noise robustness or the later MuZero system.
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