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

In animal decision-making research, reinforcement learning (RL) and working memory (WM) are regarded as two fundamental cognitive mechanisms, corresponding respectively to the accumulation of reward-based experience and the rapid utilization of recent information. This study focused on the decision-making behavior of pigeons in low- and high-difficulty probabilistic choice tasks. Based on behavioral data from five pigeons across two types of tasks, we constructed three computational models: a value-updating Rescorla-Wagner (RW) model, a limited-capacity WM model and a dual-system Rescorla-Wagner and working memory (RWWM) model with dynamic weighting. These models were used to investigate the cognitive mechanisms underlying decision making and their dynamic characteristics under varying task demands. The results revealed that pigeons continually adjusted their learning strategies in dynamic environments: WM exerted a stronger influence during the early stages of learning, facilitating rapid adaptation to changing contingencies, while RL became increasingly dominant in later stages or in more complex tasks, supporting the gradual accumulation of long-term value. Further analyses showed that in low-difficulty tasks, pigeons quickly and stably selected the option associated with the highest reward probability, consistent with predictions from the RW model. In contrast, in high-difficulty tasks, some individuals exhibited recent reward-sensitive behavior patterns more aligned with WM-based mechanisms. This study provides both computational and empirical evidence for understanding how animals flexibly deploy cognitive strategies under different learning contexts.