Value generalization, not normalization, drives learning in structure-based grouped contexts.

Area of Science:

• Cognitive Science
• Neuroscience
• Decision Science

Background:

• Value learning typically studied at the individual object level.
• Dominant account, value normalization, suggests context-relative encoding, leading to contrastive biases.
• Empirical support for normalization is limited to proximity-based grouping.

Purpose of the Study:

• Investigate context-dependent value learning in structured environments.
• Test value generalization as an alternative to value normalization.
• Examine how conceptual or relational structure influences value learning.

Main Methods:

• Five experiments using modified multi-armed bandit tasks.
• Participants inferred group structure through trial-and-error learning.
• Computational modeling compared value generalization and normalization models.

Main Results:

• Participants overestimated values of below-average items in high-mean groups.
• Assimilative bias persisted across various grouping types and reward framing.
• Value generalization models significantly outperformed normalization and item-level models.

Conclusions:

• Findings challenge the dominance of value normalization in structured environments.
• Value generalization is identified as an efficient, albeit imperfect, strategy for context-dependent value learning.
• Value learning in natural environments is influenced by inferred group structures.