Inference principles are postulated within statistics, they are not usually derived from any underlying physical constraints on real world observers. An exception to this rule is that in the context of partially observable information engines decision making can be based solely on physical arguments. An inference principle can be derived from minimization of the lower bound on average dissipation [Phys. Rev. Lett., 124(5), 050601], which is achievable with a quasi-static process. Thermodynamically rational decision strategies can be computed algorithmically with the resulting approach. Here, we use this to study an example of binary decision making under uncertainty that is very simple, yet just interesting enough to be non-trivial: observations are either entirely uninformative, or they carry complete certainty about the variable that needs to be known for successful energy harvesting. Solutions found algorithmically can be expressed in terms of parameterized soft partitions of the observable space. This allows for their interpretation, as well as for the analytical calculation of all quantities that characterize the decision problem and the thermodynamically rational strategies.

翻译：统计中的推断原则通常是假设性的，而非基于现实世界观测者所受的物理约束推导而来。然而，在部分可观测信息引擎的背景下，决策可完全依据物理论证进行。通过最小化平均耗散的下界[Phys. Rev. Lett., 124(5), 050601]（该下界可通过准静态过程实现），可推导出推断原则。基于该方法，可通过算法计算热力学理性决策策略。本文将其应用于一个简单但足够复杂的二元决策问题（在不确定性条件下）：观测要么完全无信息，要么为成功获取能量所需变量提供绝对确定性。通过算法求解得到的方案可表示为观测空间的参数化软划分，这不仅便于解释，还能解析计算表征决策问题及热力学理性策略的所有关键量。