ISSN 1671-3710
CN 11-4766/R

心理科学进展 ›› 2021, Vol. 29 ›› Issue (1): 178-189.doi: 10.3724/SP.J.1042.2021.00178

• 研究前沿 • 上一篇    下一篇


高青林, 周媛()   

  1. 中国科学院心理研究所行为科学重点实验室, 北京 100101
    中国科学院大学心理学系, 北京 100049
  • 收稿日期:2020-04-18 出版日期:2021-01-15 发布日期:2020-11-23
  • 通讯作者: 周媛
  • 基金资助:
    * 中国科学院心理研究所项目(E0CX163008)

Psychological and neural mechanisms of trust formation: A perspective from computational modeling based on the decision of investor in the trust game

GAO Qinglin, ZHOU Yuan()   

  1. Key Laboratory of Behavioral Science, Institute of Psychology, Chinese Academy of Sciences, Beijing 100101, China
    Department of Psychology, University of Chinese Academy of Sciences, Beijing, 100049, China
  • Received:2020-04-18 Online:2021-01-15 Published:2020-11-23
  • Contact: ZHOU Yuan


人际信任渗透在社会交互的各个方面, 是促进和维持合作的重要基石。以往研究者借助信任博弈范式, 主要探讨了人际信任的理论模型、生物基础和影响因素等方面。近年来, 研究者开始将计算模型应用于信任博弈的数据分析中, 深入挖掘人际信任行为背后的心理机制, 将计算模型与神经影像技术结合, 加深对信任行为背后脑机制的理解。目前将计算模型应用于信任博弈范式中的研究主要针对“信任是如何形成的”这一科学问题, 未来要进一步发展计算模型方法, 结合非侵入性脑刺激技术, 应用于精神疾病人群中, 以深入理解正常和异常信任形成的心理和神经机制。

关键词: 人际信任, 信任博弈, 计算模型, 功能磁共振成像


Interpersonal trust is the building block of corporation and resource exchange in the human society. The Trust Game is widely-used paradigm to study interpersonal trust due to its high ecological validity. The current review first introduces the TG and major paradigm variations. The aspect of paradigm variation that we especially highlight is the repeated TG (rTG), which captures the learning, reasoning and strategy updating processes compared to the single-shot TG. We then review current theoretical and empirical progresses in the computational modeling of the rTG. We specifically focus on the reinforcement learning (RL) model and the Bayesian model. The RL model essentially assumes individuals constantly update expected utility values associated with each decision options based on prediction errors, which is the difference between expected utilities and decision outcomes. The individual differences in learning capabilities are reflected in differences in learning rates. The Bayesian model essentially assumes that individuals constantly update prior beliefs about the environment based on partial observations to form posterior beliefs. The individual differences in learning capabilities are reflected in differences in the depth of theory of mind. We then reviewed behavioral studies of computational modeling of rTG. Using the RL model, previous studies have consistently shown reputation information about the trustee can not only influence initial impression about the trustworthiness level of the trustee, but also the rate of dynamic learning of feedback signal. Previous studies have formalized the Bayesian model in different ways and developed different algorithms to derive optimal behavioral solutions. These include (1) the trust level model which depicts the depth of theory of mind and the corresponding Nash-equilibrium solution, (2) the interactional partial observable Marcov Decision Process (IPOMDP) which depicts the depth of theory of mind, planning steps and greediness, and the corresponding partial observable Monte-Carlo algorithm, and (3) the subjective Bayesian model and the corresponding free-energy minimization algorithm for behavior inversion. We then review previous studies which integrate computational modeling with neuro-imaging techniques to uncover neurological bases of trust formation. The underlying logic is to extract key internal processes (e.g. the updating of prediction errors, the parameter estimation of learning rates) which are not directly observable via traditional statistics, and then establish their correlations with BOLD signals. Within the RL model, previous studies have found the updates of prediction errors were associated with the activation of striatum and anterior cingulate cortex, and the incorporation of reputation information in the trust decision process were associated with the functional interaction between the prefrontal cortex and the striatum. Within the Bayesian model, previous studies have found individuals with different depth of theory of mind primarily differed in the activation of striatum and the temporal parietal conjunction area, and that inference about intentions and adjusting behaviors based on feedback involved two independent neural systems. Lastly, we propose several prospective future directions. First, future researchers could apply the existing formalization of the Bayesian model to the rTG and apply the new type of computational modeling technique, the multilevel Gaussian filter model, which integrates both the Bayesian and the reinforcement learning component, to study the dynamic process of trust formation. Second, future researchers could incorporate non-invasive brain stimulation techniques to study the causal relationships between brain activities and behaviors. Third, computational model is an effective tool to study the impairments of trust formation among clinical samples, by comparing model formalization and parameter estimations between the clinical and healthy population. 

Key words: interpersonal trust, trust game, computational modeling, functional magnetic resonance imaging