EAGER/Collaborative Research: Challenging the Cognitive-Control Divide

这个探索性研究（EAGER）合作研究项目是由一位机器人和控制理论专家和一位实验和计算运动神经科学专家共同完成的。它是认知科学、实验心理学和控制工程的桥梁。这项工作的智力前提是，人类认知的定量理论可以建立在人类运动功能的有限案例之上。这一前提为控制相关认知的全面定量理论的发展奠定了基础。研究结果将为复杂运动控制系统的人性化设计提供宝贵的工具。基于这些基本目标的控制策略将更容易被人类操作员直观地理解，包括对即将发生的故障的预测。将结果扩展到运动控制之外，提供了一类新的知识处理系统，能够与人类进行更自然的交互。该项目的目标是阐明和测试定量的、与控制相关的人类认知理论，以解决认知科学和控制理论之间日益增长的分歧。核心假设是，认知功能来自用于运动控制的神经结构，并受其约束。复杂的运动动作由有限的动态原语“库”组成，定义为吸引子（例如不动点，极限环等）。该项目假设，类似的动态原语构成了认知过程的基础，并且可以通过重新利用运动行为中发现的动态原语来获得定量细节，特别是在操纵复杂物体（如运动和认知功能之间的联系可能最强的工具）时。该项目基于一系列新颖的实验：由不同的人类参与者物理操作复杂的动态对象生成一系列数据。替代数据集是通过计算机模拟相同目标的运动来生成的，以最小化施加的均方力。在模拟的力和运动时间序列中加入低通滤波后的零均值高斯白噪声产生的随机波动，其幅度与人类表现的波动相当。在不被告知起源的情况下，第二组受试者将结果作为演化的抽象时间序列呈现给他们，并要求他们预测结果。随后，他们被要求根据他们的经验为抽象系统生成控制输入，以完成指定的任务。根据这一假设，受试者将比合成系统更成功地预测人类控制系统的结果，并将产生更接近人类控制系统输入的控制输入。

This EArly-concept Grant for Exploratory Research (EAGER) collaborative research project is between an expert in robotics and control theory and an expert in experimental and computational motor neuroscience. It bridges cognitive science, experimental psychology and control engineering. The intellectual premise of the work is that a quantitative theory of human cognition may be built on top of limiting cases of human motor function. This premise lays the foundation for the development of a comprehensive quantitative theory of control-relevant cognition. The result will be an invaluable tool for the human-friendly design of complex motion control systems. Control strategies based on these fundamental objects would be more intuitively understandable by human operators, including prediction of impending failure. Extension of the results beyond motion control provide a new class of knowledge-processing systems capable of more natural interactions with humans. The objective of this project is to articulate and test a quantitative, control-relevant theory of human cognition, to address a growing divide between cognitive science and control theory. The core hypothesis is that cognitive functions emerged from and are constrained by neural structures used for motor control. Complex motor actions are composed from a limited "library" of dynamic primitives, defined as attractors (e.g. fixed points, limit cycles, etc.). The project postulates that a similar composition of dynamic primitives underlies cognitive processes and that quantitative details may be obtained by re-purposing dynamic primitives found in motor behavior, especially in the manipulation of complex objects such as tools where the link between motor and cognitive function may be strongest. The project is based on a novel series of experiments: A data series is generated by various human participants physically manipulating a complex dynamic object. Alternative data sets are generated by computer simulation of movements to the same targets that minimize mean-squared applied force. Random fluctuations generated by low-pass filtered zero-mean Gaussian white noise and of magnitude comparable to the fluctuations in human performance are added to the simulated force and motion time-series. Without being told the origin, a second set of subjects are presented with the results as evolving abstract time-series and asked to predict their outcome. Subsequently, they are asked to generate a control input for the abstract system, based on their experience, to accomplish a specified task. According to the hypothesis, the subjects will more successfully predict the outcome of human-controlled systems than the synthetic systems, and will generate control inputs that more closely match the human-controlled system inputs.