基于学习的多目标脑决策模型研究

摘要

Decision-making with multiple choices plays a key role on cognition.In this work,we extended a network model of Frontal eye field to a learning based model,and trained it to complete a cognitive task:non-choice taskthen used the simulation results to explain the cognitive process of multiplechoice decision-making.Afrter thousands of trainings,the network model changed from selecting target randomly into choosing the largest rewardrelative decision.In the training process,the sequence of multiplechoice decision was related to the reward gradient.In addition,the reward differences between distinct decisions played an important role on the learning speed of the network model,making the model exhibit two learning phases:the fast learning phase and the slow learning phase.%多目标决策在大脑的认知功能中起着关键的作用.在本研究中,我们将一个额叶视区网络模型扩展为一个基于学习的模型,并训练这个模型使其完成一个认知决策任务——non-choice任务,然后用模拟结果解释大脑在进行多目标选择时的认知过程.经过上千次训练后,额叶视区模型从随机选择决策目标转变为选择与最大奖励相关联的决策.在训练过程中,模型的多目标决策顺序也与目标关联的奖励梯度相关.此外,改变不同决策间的奖励差对模型的决策速度有重要的影响,可以使模型进入两种学习阶段:快速学习阶段和慢速学习阶段.