论文阅读：Towards Automatic Testing of Reference Point Based Interactive Methods

本文介绍: 本篇论文所提出的虚拟决策者Artificial Decision Maker替代人类DM向交互式算法提供参考点，但是参考点的生成方式并没有涉及到人类DM的经验，本文是通过预设的目标函数期望水平ADM aspirations point(asp)与当前Pareto解集中各目标函数值来确定的。

Towards Automatic Testing of Reference Point Based Interactive Methods

作者：Vesa Ojalehto、Dmitry Podkopaev、Kaisa Miettinen
期刊：PPSN、2016
DOI：10.1007/978-3-319-45823-645

内容简介

为了了解优化算法的优缺点，有必要了解不同类型的测试问题、定义良好的性能指标和分析工具。这样的工具广泛用于测试进化多目标优化算法。

据我们所知，目前还没有基于参考点来传递偏好信息的交互式多目标优化方法的性能分析工具。这些工具的主要障碍是人类决策者参与交互式解决方案进程，这使得交互式方法的性能取决于人类使用它们的性能。在这项研究中，我们的目标是建立一个测试框架，在这个框架中，人工决策者被替换为虚拟决策者，并且允许在一个可控的环境中重复测试交互式方法。

内容摘录

We propose to employ an artificial DM to replace the real DM. Our concept of an artificial DM and its interaction with an interactive method comprises the following three components：

Steady part: the complexity of knowledge possessed by the DM and related to solving the considered class of problems which does not change during the solution process. This includes accumulated experience and the core preferences which do not change in time.

优化过程中不随时间变化的内容(经验、偏好、问题类型、参数)

Current context: the current situation as perceived by the DM, which may change in time. This includes: the knowledge about the problem accumulated by the DM during the solution process, level of tiredness which can affect concentration, and the probability of making mistakes.

优化过程中随着时间会发生变化的内容(DM积累的经验、影响注意力的疲劳程度、失误率)

Preference information: the method-specific information expressed by the DM during the solution process to guide the method toward solutions that are more preferred by the DM.

The artificial DM should be defined by the steady part which does not change in time, a mechanism of representing and updating the current context as the solution process continues, and the mechanism of generating the preference information based on the steady part and the current context. By varying the parameters of the steady part, one can obtain different artificial DMs for conducting multiple experiments

虚拟决策者应该由不随时间变化的稳定部分来定义，随着求解进程的继续，表达和更新当前上下文的机制、基于稳定部分和当前上下文生成偏好信息的机制。通过改变稳态部分的参数，可以得到不同的虚拟决策者进行多次实验。