Using recurrent neural network to estimate irreducible stochasticity in human choice behavior

Area of Science:

• Cognitive science
• Computational neuroscience
• Decision-making science

Background:

• Theoretical computational models explain cognitive processes but vary in predictive accuracy across individuals.
• A significant predictive gap exists for some participants, necessitating an investigation into its source.

Purpose of the Study:

• To utilize theory-independent recurrent neural networks (RNNs) to differentiate between noisy decision-making and theoretical model misspecification.
• To assess whether individual differences in behavioral data predictability stem from stochasticity or model inadequacy.

Main Methods:

• Computer simulations using reinforcement learning to train RNNs on simulated agents with controlled noise levels.
• Analysis of prediction performance and RNN training epochs (early stopping) to quantify stochasticity.
• Application of the RNN approach to an empirical dataset of human participants performing a two-step task, comparing low and high IQ groups.

Main Results:

• RNNs successfully identified model misspecification in simulated agents, with prediction performance and early stopping indicating stochasticity levels.
• In the empirical dataset, RNN analysis suggested similar levels of model misspecification across all participants.
• The predictive gap and RNN early stopping indicated that lower predictability in low IQ participants was primarily due to increased decision-making noise.

Conclusions:

• Recurrent neural networks offer a method to diagnose the source of predictive gaps in computational models of cognition.
• Low IQ participants exhibit noisier decision-making rather than greater misspecification by established theoretical models.
• The findings highlight the utility of data-driven, theory-independent models in complementing traditional theoretical approaches in decision science.