Interaction between dynamic reinforcement learning and working memory of pigeon: a comparative modeling study

Area of Science:

• Cognitive Neuroscience
• Animal Behavior
• Computational Modeling

Background:

• Reinforcement learning (RL) and working memory (WM) are key cognitive mechanisms in animal decision-making.
• RL involves accumulating reward experience, while WM utilizes recent information.
• Understanding their interplay is crucial for deciphering adaptive behavior.

Purpose of the Study:

• To investigate the cognitive mechanisms underlying pigeon decision-making in probabilistic choice tasks.
• To examine how reinforcement learning and working memory are dynamically deployed under varying task demands.
• To model and differentiate the contributions of RL and WM in decision strategies.

Main Methods:

• Behavioral experiments with five pigeons across low- and high-difficulty probabilistic choice tasks.
• Construction and comparison of three computational models: Rescorla-Wagner (RW), Working Memory (WM), and a dual RWWM model.
• Analysis of behavioral data to assess model fit and infer cognitive strategies.

Main Results:

• Pigeons dynamically adjusted learning strategies, with WM dominating early learning and RL dominating later stages or complex tasks.
• In low-difficulty tasks, pigeons favored the RW model, showing stable selection of high-reward options.
• In high-difficulty tasks, some pigeons displayed WM-like behavior, sensitive to recent rewards.

Conclusions:

• Pigeons exhibit cognitive flexibility, dynamically weighting working memory and reinforcement learning based on task context.
• Working memory aids rapid adaptation, while reinforcement learning supports long-term value accumulation.
• This study provides computational and empirical evidence for flexible strategy deployment in animal decision-making.