Contextual influence on confidence judgments in human reinforcement learning.

Area of Science:

• Cognitive Neuroscience
• Decision Science
• Reinforcement Learning

Background:

• Accurate metacognition, or confidence judgment, is crucial for decision-making and strategy adaptation.
• Metacognitive processes are known to be susceptible to various biases.
• Understanding these biases is key to optimizing decision-making strategies.

Purpose of the Study:

• To investigate the impact of outcome valence (gains vs. losses) on confidence judgments during learning.
• To explore the underlying computational mechanisms driving confidence biases in different learning contexts.
• To examine the functional consequences of these biases, particularly in dynamic environments.

Main Methods:

• Two experiments involving trial-and-error learning of stimulus-outcome associations.
• Behavioral data collection on choices and confidence ratings.
• Computational modeling to analyze confidence biases and context-value estimation.

Main Results:

• Participants exhibited higher confidence when learning to gain rewards compared to avoiding losses, despite equivalent performance and task difficulty.
• Computational models identified context-value as the driver of this confidence bias.
• The study revealed domain-general effects of context-value on confidence in reinforcement learning.

Conclusions:

• Confidence judgments are asymmetrically influenced by the framing of outcomes as gains or losses.
• The context-value, an estimate of average expected value, significantly biases metacognitive confidence.
• These findings highlight potential asymmetries in learning to seek gains versus avoid losses, with implications for behavioral flexibility in changing environments.