Simulating probability learning and probabilistic reversal learning using the attention-gated reinforcement learning (AGREL) model

In a probability learning task, participants estimate the probabilistic reward contingencies, and this task has been used extensively to study instrumental conditioning with partial reinforcement. In the probabilistic reversal learning task, the probabilistic reward contingencies are reversed between options in the middle of the experiment to measure how well people adapt to new contingency situations. In this work, we used the attention-gated reinforcement learning (AGREL) model (Roelfsema & Van Ooyen, 2005) to simulate how people learn the probabilistic relationship between stimulus-reward pairs in probability and reversal learning tasks. AGREL algorithm put forward two important aspects of a learning phenomenon together in a neural network scheme: (1) the effect of unexpected outcomes on learning and (2) the effect of top-down (selective) attention on updating weights. Contrary to its importance in the learning literature, AGREL has not yet been tested with these well known learning tasks. The results of the first simulation showed that in a binary choice probability learning experiment an AGREL model can simulate different learning strategies, such as probability matching and maximizing. Secondly, we simulated a probabilistic reversal learning experiment with the same AGREL model, and the results showed that the AGREL model dynamically adapted to new contingency situations. Furthermore, we also evaluated effects of learning rate on the model's adaption to reversal contingency by plotting the interphase dynamics. These results showed that AGREL model simulates the traditional findings observed in probability and reversal learning experiments, and it can be further developed to understand the role of dopamine in learning and it can be used in model-based fMRI research.