Lebiere Christian, Pirolli Peter, Johnson Matthew, Martin Michael, Morrison Donald
Department of Psychology, Carnegie Mellon University.
Institute for Human & Machine Cognition, Pensacola.
Top Cogn Sci. 2025 Apr;17(2):268-290. doi: 10.1111/tops.12773. Epub 2024 Dec 1.
Some of the required characteristics for a true machine theory of mind (MToM) include the ability to (1) reproduce the full diversity of human thought and behavior, (2) develop a personalized model of an individual with very limited data, and (3) provide an explanation for behavioral predictions grounded in the cognitive processes of the individual. We propose that a certain class of cognitive models provide an approach that is well suited to meeting those requirements. Being grounded in a mechanistic framework like a cognitive architecture such as ACT-R naturally fulfills the third requirement by mapping behavior to cognitive mechanisms. Exploiting a modeling paradigm such as instance-based learning accounts for the first requirement by reflecting variations in individual experience into a diversity of behavior. Mechanisms such as knowledge tracing and model tracing allow a specific run of the cognitive model to be aligned with a given individual behavior trace, fulfilling the second requirement. We illustrate these principles with a cognitive model of decision-making in a search and rescue task in the Minecraft simulation environment. We demonstrate that cognitive models personalized to individual human players can provide the MToM capability to optimize artificial intelligence agents by diagnosing the underlying causes of observed human behavior, projecting the future effects of potential interventions, and managing the adaptive process of shaping human behavior. Examples of the inputs provided by such analytic cognitive agents include predictions of cognitive load, probability of error, estimates of player self-efficacy, and trust calibration. Finally, we discuss implications for future research and applications to collective human-machine intelligence.
真正的心智机器理论（MToM）所需的一些特性包括能够：（1）再现人类思想和行为的全部多样性；（2）在数据非常有限的情况下为个体建立个性化模型；（3）为基于个体认知过程的行为预测提供解释。我们提出，某一类认知模型提供了一种非常适合满足这些要求的方法。基于诸如ACT-R这样的认知架构的机械框架，通过将行为映射到认知机制，自然地满足了第三个要求。利用诸如基于实例的学习这样的建模范式，通过将个体经验的变化反映到行为的多样性中，满足了第一个要求。诸如知识追踪和模型追踪等机制允许认知模型的特定运行与给定的个体行为轨迹对齐，满足了第二个要求。我们在Minecraft模拟环境中的搜索和救援任务的决策认知模型中说明了这些原则。我们证明，针对个体人类玩家个性化的认知模型可以通过诊断观察到的人类行为的潜在原因、预测潜在干预的未来影响以及管理塑造人类行为的自适应过程，提供优化人工智能代理的MToM能力。这种分析性认知代理提供的输入示例包括认知负荷预测、错误概率、玩家自我效能估计和信任校准。最后，我们讨论了对未来研究以及集体人机智能应用的启示。
