EAGER: Enabling Taskability: Research on Teaching Computers new Tasks

这项研究探索了认知主体“可任务性”的一条新的研究路线。任务能力是指代理接受人类的高级、面向任务的指令，并将这些面向目标的指令转换为适当分解的感知、推理或行动计划的能力。到目前为止，关于这一主题的研究很少，甚至没有。这项研究以足够正式的方式定义了可任务性，以便进行科学研究，建立了评估可任务代理所需的一些初始条件，并提出了满足这些要求的理论和原型代理。具体地说，本研究的四项研究活动包括：1.回顾、分析和综合多个学科的先前工作，为这一领域的认知系统研究奠定基础；2)分析不同类型的任务及其结构，以及相关类型的知识，这些知识必须由任务代理学习；3)研究人们如何使用任务代理进行指导，执行用户研究以确定所需的任务知识和代理能力；4)扩展现有的认知主体能力，以支持上述研究中观察到的行为，从而产生一个可以与理论一起评估和共同开发的软件系统。在认知系统和代理的研究中，目前的技术水平通常允许人类操作者以与代理的特定设计相对应的较低级别的动作基元的形式向代理发出指示，或者受到高阶动作语言(甚至是视觉曲目)的约束，该语言被设计为足以作为相同类型的代理特定行为的某些序列的简写。可任务性要求代理能够解释来自用户的更丰富、更少约束的指令集，并且尽管缺乏预编译的任务分解指令，但能够动态地表示要执行的任务的准确表示，甚至通过这种交互来学习新的任务。可执行任务的代理将从根本上改变人类在广泛的学科和环境中与智能代理和机器人系统交互的方式，导致与功能更强大的汽车、电话和几乎任何包含计算机的设备进行更丰富的交互，所有这些设备都可以响应人类交互，而不需要求助于预先编程的交互模式。这反过来又承诺在医疗保健、工业、政府和家庭自动化等一系列应用领域取得变革性的进展。

This study investigates a new line of research into the "taskability" of cognitive agents. Taskability is the ability of an agent to accept high level, task-oriented instructions from a human and translate those goal-oriented directives into a suitably decomposed plan of sensing, reasoning, or action. Little to no research has been conducted on this subject to date. This research defines taskability in a manner sufficiently formal to allow for scientific research, establishes some of the initial conditions needed to evaluate taskable agents, and advances theories and prototype agents that meet those requirements. Specifically, the four research activities being undertaken in this study include: 1. A review, analysis, and synthesis of prior work from multiple disciplines that that can lay groundwork for focused cognitive systems research in this area; 2) An analysis of the different types of tasks and their structure, as well as the associated types of knowledge that must be learned by taskable agents; 3) Research on how people use instruction with taskable agents for such activities, performing user studies to determine the required task knowledge and agent capabilities; and 4) An extension of current cognitive agent capabilities to support the behaviors observed in the aforementioned studies, resulting in a software system that can be evaluated and co-developed with theory. The current state of the art in research for cognitive systems and agents in general allows human handlers to issue directions to agents either in terms of lower level action primitives that correspond to the agent's particular design, or constrained by a higher-order action language (or even visual repertoire) that is engineered to suffice as a shorthand for some sequence of the same kinds of agent-specific behaviors. Taskability requires that an agent be able to interpret a richer, less constrained set of instructions from a user and, despite a lack of precompiled task decomposition instructions, dynamically formulate an accurate representation of the task to be performed, and even learn new tasks via this sort of interaction. Taskable agents would fundamentally change the way humans interact with intelligent agents and robotic systems in a broad range of disciplines and environments, leading to richer interactions with more capable cars, phones, and almost any device containing a computer, all of which might respond to human interaction without resort to preprogrammed interaction modes. This in turn promises transformational advances in a range of application domains such as health care, industry, government, and home automation.