RI: Small: The Dynamics of Information Flow in Embodied Cognitive Systems

该奖项是根据2009年美国复苏和再投资法案(公法111-5)资助的。行为和脑科学领域越来越多的人认识到，智能代理人的具身和位置在他们的行为中起着至关重要的作用。然而，理解代理人的神经系统、其身体和环境之间的复杂相互作用仍然是一个重大的开放挑战。本项目侧重于用数学方法分析情境认知主体和具身认知主体模型中的信息流。具体地说，研究人员将(1)开发新的信息论工具来表征整个系统中的信息流，(2)在进化模型脑-体-环境系统上测试和改进这些方法，以及(3)使用这些技术和模型来探索具体化和情境性对于认知主体的独特优势。新的分析技术建立在特定时间点随机变量之间的条件互信息的概念上。例如，以时间t处的刺激特征和状态变量之间的互信息为条件，以状态变量包含的关于时间t-1处的刺激特征的信息为条件，允许计算信息增益的测量。可以使用类似的衡量标准来计算信息损失或保留率。这一基本方法将在几个不同的方向上扩展。建立在传递熵概念基础上的信息分解将允许根据定向信息流来表征系统组件之间的相互作用。信息回溯将通过及时回溯系统的特定信息配置来确定产生这些交互的流，从而提供对这些交互的结构的精细描述。将使用特定的信息光谱来探索系统组件所采用的特定值之间的信息关系，从而能够在更详细的水平上探测其相互作用的结构。这些方法将被应用于关系分类、参照交流和视觉引导行为的进化模型，允许探索以下问题：具体化代理如何提取、存储和抑制信息？具体化代理如何集成有关多个功能的信息？信息属性和动力学属性之间的关系是什么？最后，这些技术将被用来表征具身和情境代理的能力，包括信息自组织、信息卸载和具身信息传递。这些分析方法不仅有望应用于脑-体-环境系统，还将应用于其他复杂的生物和社会网络。

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).There is a growing realization within the behavioral and brain sciences that the embodiment and situatedness of intelligent agents plays an essential role in their behavior. However, it is still a significant open challenge to understand the complex interactions between an agent's nervous system, its body and its environment. This project focuses on mathematical methods for analyzing the flow of information in models of situated and embodied cognitive agents. Specifically, the investigator will (1) develop new information-theoretic tools that characterize the flow of information over time throughout the system, (2) test and refine these methods on evolved model brain-body-environment systems, and (3) use these techniques and models to explore the unique advantages of embodiment and situatedness for a cognitive agent.The new analysis techniques build on the notion of conditional mutual information between random variables at specific points in time. For example, conditioning the mutual information between a stimulus feature and a state variable at time t on the information that state variable contains about the stimulus feature at time t-1 allows one to compute a measure of information gain. Similar measures can be used to compute the information loss or retention. This basic approach will be extended in several different directions. Information factoring, which builds on the notion of transfer entropy, will allow the interactions between system components to be characterized in terms of directional information flow. Information backtracking will offer a refined portrait of the structure of these interactions by tracing backwards in time from particular informational configurations of the system to determine the flows that produce them. Specific information spectra will be used to explore the informational relationships between particular values taken on by components of the system, allowing the structure of their interactions to be probed at a finer level of detail. These methods will be applied to evolved models of relational categorization, referential communication, and visually guided behavior, allowing the following questions to be explored: How do embodied agents extract, store, and suppress information? How do embodied agents integrate information about multiple features? What is the relationship between informational and dynamical properties? Finally, these techniques will be used to characterize the capabilities of embodied and situated agents, including information self-structuring, information offloading, and embodied information transfer. These analysis methods are expected to have applications not only to brain-body-environment systems, but also to other complex biological and social networks.