The Cost of Decision-Making: Unraveling the Interplay of Memory, Reinforcement Learning, and Neural Connectivity in Decision-Making

Abstract

This dissertation reviewed significant theoretical work and empirical findings in the field of decision-making, and posed critical questions spanning various aspects of decision-making, such as cognitive control, motivation, individual differences, memory, and computational modeling. To address these questions, I present three computational works, aiming to improve understanding of decision-making processes within the ACT-R cognitive framework. The first computational work outlines a novel framework for cognitive effort allocation, proposing that optimal cognitive control is based on a balance of cognitive effort and benefits. The second piece of research identifies two decision-making strategies: memory-based and reinforcement-based, and uncovers how individuals' preferences for one strategy over another are predicted by their resting-state brain connectivity. By employing computational cognitive models and ML models, this study demonstrated that stronger connectivity in frontoparietal and memory retrieval regions corresponds to a preference for memory-based strategies, while stronger connectivity in sensorimotor, cingulate, and basal ganglia regions predicts a preference for reinforcement-based strategies. This work highlights the adaptive nature of human decision-making in both behavioral and neurofunctional data. Lastly, the third piece of computational research reconciles two primary learning approaches in complex decision-making tasks, introducing a cognitively plausible model that combines memory retrieval and model-based reinforcement learning (MB-RL). This integrated model connects established findings in reinforcement learning and offers insights into the balance between memory and reinforcement learning in decision-making, paving the way for advancements in artificial intelligence and decision-making modeling.