在研究人類如何做決定的這門學問中，學術領域的疆界不斷的在被打破。從1970年代的心理學，以Kahneman和Tversky為首，正式打入經濟學領域，到1990年代末，在Paul Glimcher, Colin Camerer等人帶動下，開始發展的神經經濟學，甚至在最近兩年，Deep Mind和神經科學家Nao Uchida合作，在大腦神經細胞內找到人工智慧新興演算法的生理證據。到底誰是研究決策的「專家」? 從來沒有人能說的準。因此，這門進階課的初衷，是特別為來自不同背景學員所規劃的一門課，集合四位頂尖的新進神經科學家來分享他們的研究，將從人類和動物如何從經驗中學習、偏好如何形成、決策偏誤的來源、到認知策略如何影響行為等角度，來剖析決策行為背後迷人的神經機制。你可以從中獲得神經科學家們，如何透過實驗設計、如何利用數學模型和演算法、如何巧妙使用電生理和神經造影等技術，來了解大腦與行為間的奧秘。相信你們能在這個集合神經、經濟、心理、資訊和工程專才的跨領域研究中，有豐富的收穫！

神經經濟學（Neuroeconomics）認為當大腦進行選擇時，會為各個選項打分數，進行比較後做出選擇。然而，「酬賞價值」（reward value）如何在大腦中產生，卻是決策研究中極少被探究的根本問題。我們團隊利用成分經過特殊設計的飲料作為酬賞，研究獼猴在選擇不同飲料時，如何感知並整合飲料的感覺和營養特性，形成對酬賞的主觀價值。根據此概念，我們更發展出仿生的強化學習（reinforcement learning）演算法，能在環境發生變動時，根據主觀的營養偏好，從經驗中學習，並且修正選擇。因為營養攝取對動物生存與物種延續至關重要，透過結合生態理論的獼猴營養決策與神經紀錄實驗，我們能進一步研究大腦的獎賞系統如何受到營養限制的形塑，演化出調控複雜的認知與決策行為的神經機制。

In uncertain environments, learning is driven by prediction errors that capture the difference between people’s beliefs and observations. Past studies on reinforcement learning focused on the cognitive mechanism but paid less attention on the affective mechanism. In this talk, I will discuss both cognitive and affective mechanisms in learning and demonstrate how prediction errors influence behaviors and mood. First, I will describe how to use prediction errors to update beliefs. I will start with learning in stable environments and extend to dynamic environments. In dynamic environments, the state of the environment can change. People should adjust the speed of their learning adaptively. They should maintain their beliefs when the state is stable, and quickly update their beliefs when the state changes. In several studies, I showed the computational mechanism to adjust learning. Using fMRI, I identified how the communication between frontoparietal network and other brain networks is related to the adjustment of learning. Using machine learning techniques, I further distinguished the functional roles of frontal and parietal regions. Posterior parietal cortex encoded prediction errors in a task-dependent manner while frontal cortex predicted the subsequent behavioral change. Second, I will focus on the affective mechanism during learning and demonstrate how to measure mood dynamics. Using computational models, I demonstrated how prediction errors drive the changes of mood during learning. In sum, this talk will discuss the cognitive and affective mechanisms in reinforcement learning and show how different brain regions contribute to learning.