A reinforcement learning diffusion decision model for value-based decisions.

Area of Science:

• Cognitive Neuroscience
• Computational Psychology
• Behavioral Economics

Background:

• Value-based decision-making models traditionally focus on choice outcomes.
• Sequential sampling models, like the diffusion decision model (DDM), incorporate response times but often neglect learning.
• Integrating reinforcement learning with decision models is crucial for understanding dynamic choice behavior.

Purpose of the Study:

• To propose and validate a novel Reinforcement Learning Diffusion Decision Model (RLDDM).
• To investigate how learning from rewards influences decision-making dynamics.
• To bridge the gap between reinforcement learning and sequential sampling models in decision research.

Main Methods:

• Development of the combined Reinforcement Learning Diffusion Decision Model (RLDDM).
• Testing the RLDDM on a learning task with options varying in value difference and overall value.
• Comparison of RLDDM performance against standard diffusion decision models (DDMs) and reinforcement learning models.

Main Results:

• Participants improved accuracy and speed with learning.
• Faster and more accurate responses were observed for options with greater value dissimilarity.
• Decisions were faster when presented with more attractive option pairs.
• The RLDDM successfully accounted for these observed behavioral effects.

Conclusions:

• The proposed RLDDM provides a unified framework for understanding value-based decisions with learning.
• RLDDM offers a superior account of decision dynamics compared to existing models.
• This work advances computational models of decision-making by integrating learning and temporal dynamics.