Social Prediction Error Gating for Adaptive Social Learning: A Computational Model of Individual and Social Information Integration

This paper presents Social Prediction Error Gating (SPEG), a computational reinforcement-learning model designed to combine individual experience with social information based on estimated partner reliability and uncertainty. The proposed model is evaluated through simulations of patch-selection foraging in stable and changing environments, under different levels of individual and social information reliability. SPEG is compared with partner-aware Q-learning, a dual-learning-rate learning model, and a partner-ignoring Q-learning baseline. The results show that SPEG did not outperform partner-aware Q-learning in cumulative reward or adaptation after environmental changes. The social weighting mechanism remained close to equal weighting across conditions, indicating that it did not effectively distinguish between reliable and unreliable social information. These findings highlight the importance of designing effective reliability-estimation mechanisms and adaptive gating strategies in social reinforcement learning. The study provides a computational analysis of individual and social information integration and identifies directions for future research.