強化学習における認知バイアスと固執性―選択行動を決めているのは過去の“選択の結果”か“選択そのもの”か?―

抄録

Reinforcement learning models, which update the value related to a specific behaviour according to a reward prediction error, have been used to model the choice behaviour in organisms. Recently, the magnitude of the learning rate has been reported to be biased depending on the sign of the reward prediction error. A previous study concluded that these asymmetric learning rates reflect positivity and confirmation biases. However, another study reported that the tendency to repeat the same choice (perseverance) leads to pseudo asymmetric learning rates. Therefore, this study aimed to clarify whether asymmetric learning rates are the result of cognitive bias or perseverance by reanalysing the open data that the previous study obtained from two different types of learning tasks. To accomplish this, we evaluated multiple reinforcement learning models, including asymmetric learning rate models, perseverance models and hybrid models. The results showed that the choice data associated with positivity bias were also explained by the perseverance model with symmetric learning rates. Meanwhile, the data associated with confirmation bias were not explained by the perseverance model. These results suggest the possibility that either cognitive bias or perseverance could explain asymmetric learning rates depending on the contextual information of learning task.