关键词: 系列决策任务, 爱荷华赌博任务, 策略转换, 计算认知建模, 强化学习和启发式策略

Abstract:

Much research has been devoted to studying decision strategies in various tasks. Such research usually involved a sequence of decision trials under the same task structure to provide sufficient information for inferring the underlying decision strategies. By assuming each individual adopted a single decision strategy across all decision trials and comparing corresponding computational cognitive models in terms of their performances in fitting empirical data, such studies have revealed multiple possible decision strategies for many major decision tasks. One common drawback of such research, however, was overlooking the possibility that individuals switched their strategies along the sequence of decisions. This might lead to inappropriate conclusions regarding the decision strategies underlying specific decision tasks or misleading inferences of potential cognitive and affective differences between normal and different clinical populations based on parameter estimates from models assuming single strategies.
To address this critical issue, two studies were conducted to examine the possibility of strategy switching in the Iowa Gambling Task (IGT), an experience-based decision task with a sequence of trials aimed at mimicking real-world decisions under uncertainty. By developing a computational cognitive model that allowed for switches between reinforcement learning strategies and heuristic strategies and comparing its performance with those of single-strategy models, Study 1 showed that data from about half of the 617 healthy participants in 10 previous studies were better fitted by the strategy-switching model than three single-strategy models that performed well in previous research, that is, the WSLS, PVL2, and VPP models as exemplar models assuming heuristic, reinforcement learning, and mixed strategies, respectively. This result provided clear support for the possibility of strategy switching in the IGT.
Since strategy switching might occur with accumulating experience or fatigue and an increasing number of trials is likely to facilitate such changes, 321 participants were recruited in Study 2 to further examine whether a larger number of trials would contribute to more strategy switching in the IGT. Specifically, 160 participants performed a 100-trial IGT, whereas the other 161 participants performed a 200-trial IGT under otherwise the same task structure. It was found that data from a larger proportion of individual participants were best fitted by the strategy-switching model when the IGT involved 200 trials rather than standard 100 trials. This result provided further evidence for strategy switching in the task.
Overall, the current results suggest that strategy switching is likely to occur in a sequence of decisions under the same task structure. Consequently, in order to obtain proper understanding of the decision strategies for various decision tasks, it is necessary to consider seriously the possibility of strategy switching, especially for a long sequence of decisions. For a more refined understanding of psychological mechanisms underlying sequences of decisions, future research might further investigate various forms of strategy switching such as gradual instead of abrupt switches and task and individual factors that trigger such switches.

Key words: decision task with a sequence of trials, The Iowa Gambling Task, strategy switching, computational cognitive modeling, reinforcement learning and heuristic strategies