User Interface Softbots

这个项目的重点是ibots，通过图形用户界面与软件应用程序交互的界面代理，通过与人类用户相同的媒介。该项目将有两个主要重点。首先，PI将根据各种常见应用程序中的用户行为描述，定义用户界面中工具使用的分类，并且他将构建代理工具使用的计算模型，该模型集成了可视化例程和基于成本的用户建模工作。然后将开发一个用于智能体探索和评估的环境，包括三个不同任务领域的工具集：支持领域描述的构建和智能体开发的增量构建/测试周期；支持在执行这些任务时探索不同的控制策略；并支持对接口代理性能的评价。随着该环境的完成，PI将沿着两条并行轨道对其进行评估，作为一种允许认知模型与现成软件交互的机制，重点关注ACT-R、Soar和LICAI框架中的模型，并作为开发规划算法新问题域的测试平台。这项工作将在几个领域产生影响。它将允许认知模型研究人员在实际（非模拟）环境中评估他们的模型，这是该领域研究人员长期以来的目标。它将为规划研究人员提供类似的好处，他们也需要既真实又易于处理的规划算法域。它将产生新的接口代理和领域描述，并有机会以相对较低的开发成本测试现有代理设计对现实世界问题的充分性；该项目也有可能增加界面代理研究的可见性，允许代理在以前无法进入的领域工作。最后，它将产生一个专门用于工具使用的代理的模型和实现，这是迄今为止尚未检验的生态研究领域，它将对我们对复杂软件环境的理解做出重大贡献。

This project focuses on ibots, interface agents that interact with software applications through the graphical user interface, through the same medium as human users. The project will have two main thrusts. First, the PI will define a taxonomy of tool use in the user interface, based on descriptions of user behavior in a variety of common applications, and he will build a computational model of agent tool use that integrates work on visual routines and cost-based user modeling. An environment for agent exploration and evaluation will then be developed, to include sets of tools in three distinct task areas: support for the construction of domain descriptions and the incremental build/test cycle for agent development; support for the exploration of different control strategies in carrying out these tasks; and support for the evaluation of interface agent performance. As this environment is completed, the PI will evaluate it along two parallel tracks, as a mechanism allowing cognitive models to interact with off-the-shelf software, focusing on models in the ACT-R, Soar, and LICAI frameworks, and as a testbed for developing new problem domains for planning algorithms. This work will have an impact in several areas. It will allow cognitive modeling researchers to evaluate their models in actual (not simulated) environments, a long-standing goal for researchers in this area. It will provide a comparable benefit to planning researchers, who also need domains that are both real and tractable for their planning algorithms. It will produce novel interface agents and domain descriptions, with opportunities to test the adequacy of existing agent designs against real-world problems at a relatively low development cost; the project also has the potential to increase the visibility of interface agents research, by allowing agents to work in previously inaccessible areas. Finally, it will result in a model and implementation of agents specialized for tool use, an area of ecological research that remains unexamined to date, and which should contribute significantly to our understanding of complex software environments.