Lim Sewoong, Yoon Sangsup, Kwon Jaehyung, Kralik Jerald D, Jeong Jaeseung
Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea.
Front Comput Neurosci. 2020 Sep 11;14:65. doi: 10.3389/fncom.2020.00065. eCollection 2020.
Humans organize sequences of events into a single overall experience, and evaluate the aggregated experience as a whole, such as a generally pleasant dinner, movie, or trip. However, such evaluations are potentially computationally taxing, and so our brains must employ heuristics (i.e., approximations). For example, the peak-end rule hypothesis suggests that we average the peaks and end of a sequential event vs. integrating every moment. However, there is no general model to test viable hypotheses quantitatively. Here, we propose a general model and test among multiple specific ones, while also examining the role of working memory. The models were tested with a novel picture-rating task. We first compared averaging across entire sequences vs. the peak-end heuristic. Correlation tests indicated that averaging prevailed, with peak and end both still having significant prediction power. Given this, we developed generalized order-dependent and relative-preference-dependent models to subsume averaging, peak and end. The combined model improved the prediction power. However, based on limitations of relative-preference-including imposing a potentially arbitrary ranking among preferences-we introduced an absolute-preference-dependent model, which successfully explained the remembered utilities. Yet, because using all experiences in a sequence requires too much memory as real-world settings scale, we then tested "windowed" models, i.e., evaluation within a specified window. The windowed (absolute) preference-dependent (WP) model explained the empirical data with long sequences better than without windowing. However, because fixed-windowed models harbor their own limitations-including an inability to capture peak-event influences beyond a fixed window-we then developed discounting models. With (absolute) preference-dependence added to the discounting rate, the results showed that the discounting model reflected the actual working memory of the participants, and that the preference-dependent discounting (PD) model described different features from the WP model. Taken together, we propose a combined WP-PD model as a means by which people evaluate experiences, suggesting preference-dependent working-memory as a significant factor underlying our evaluations.
人类将一系列事件组织成一个单一的整体体验，并对整体的综合体验进行评估，比如一顿总体愉快的晚餐、一场电影或一次旅行。然而，这样的评估在计算上可能很费力，所以我们的大脑必须采用启发式方法（即近似法）。例如，峰终定律假说表明，我们会对连续事件的峰值和结尾进行平均，而不是整合每一个瞬间。然而，目前还没有一个通用模型来定量检验可行的假说。在这里，我们提出一个通用模型，并在多个具体模型中进行测试，同时也考察工作记忆的作用。这些模型通过一项新颖的图片评分任务进行测试。我们首先比较了对整个序列进行平均与峰终启发式方法。相关性测试表明平均法占主导地位，峰值和结尾仍然都具有显著的预测能力。鉴于此，我们开发了广义顺序依赖和相对偏好依赖模型，以包含平均法、峰值和结尾。组合模型提高了预测能力。然而，基于相对偏好的局限性——包括在偏好之间强加一个可能任意的排序——我们引入了一个绝对偏好依赖模型，该模型成功地解释了记忆中的效用。然而，由于在现实世界环境中，使用序列中的所有体验需要太多记忆，我们随后测试了“窗口化”模型，即在指定窗口内进行评估。窗口化（绝对）偏好依赖（WP）模型比无窗口化时能更好地解释长序列的实证数据。然而，由于固定窗口模型有其自身的局限性——包括无法捕捉固定窗口之外的峰值事件影响——我们随后开发了折扣模型。在折扣率中加入（绝对）偏好依赖性后结果表明，折扣模型反映了参与者的实际工作记忆情况，并且偏好依赖折扣（PD）模型描述了与WP模型不同的特征。综上所述，我们提出一个组合的WP - PD模型，作为人们评估体验的一种方式，表明偏好依赖的工作记忆是我们评估背后的一个重要因素。
