Abstract In the free energy principle (FEP) proposed by Friston, it is supposed that agents seek to minimize the “surprise” – the negative log (marginal) likelihood of observations (i.e., sensory stimuli) – given the agents’ current belief. This is achieved by minimizing the free energy, which provides an upper bound on the surprise. The FEP has been applied to action selection in a framework called “active inference,” where agents are supposed to select an action so that they minimize the “expected free energy” (EFE). While EFE can be decomposed into interpretable components such as epistemic value and extrinsic value, it is difficult to understand intuitively how EFE itself is directly related to “surprise” and what psychological construct is related to EFE itself (as a single quantity). To facilitate the discussion and interpretation of psychological processes underlying active inference, we introduce a computational component termed the “retrospective surprise,” which is the surprise of an observation after updating the belief given the observation itself. The predicted retrospective surprise (PRS) is mathematically derived from a special case of EFE and provides a lower bound on EFE. We illustrate the properties of EFE and PRS using examples of inference for a binary hidden cause given a binary observation. We discuss how information-seeking behavior is accounted for by EFE and PRS. Our results highlight the role of prior distribution of future observation in EFE. By setting this prior distribution to be fixed irrespective of which action is selected, the epistemic value can exert an influence on EFE, leading to information-seeking behavior.

中文翻译：

---

回顾性惊喜：用于主动推理的计算组件

摘要 在弗里斯顿提出的自由能原理 (FEP) 中，假设代理寻求最小化“惊喜”——观察（即感官刺激）的负对数（边际）可能性——给定代理当前的信念。这是通过最小化自由能来实现的，自由能提供了惊喜的上限。FEP 已应用于称为“主动推理”的框架中的动作选择，其中代理应该选择一个动作，以便它们最小化“预期自由能”（EFE）。虽然 EFE 可以分解为认知价值和外在价值等可解释成分，但很难直观地理解 EFE 本身如何与“惊喜”直接相关，以及什么心理构造与 EFE 本身相关（作为一个单一的数量）。为了促进对主动推理背后的心理过程的讨论和解释，我们引入了一个称为“回顾性惊喜”的计算组件，这是在更新给定观察本身的信念后观察的惊喜。预测的回顾性惊喜 (PRS) 是从 EFE 的一个特殊情况在数学上推导出来的，并提供了 EFE 的下限。我们使用给定二元观察的二元隐藏原因的推理示例来说明 EFE 和 PRS 的属性。我们讨论了 EFE 和 PRS 如何解释信息寻求行为。我们的结果强调了未来观察的先验分布在 EFE 中的作用。通过将此先验分布设置为固定，无论选择哪个动作，认知值都可以对 EFE 产生影响，