Explaining and Architecting Intelligent and Connected Software Systems

大型复杂软件系统的特征正在发生变化。软件变得越来越智能，增加了人类难以解释的机器学习组件。软件正变得越来越相互关联，由部分编排的交互组件集组成，这些组件创建的结构仅在运行时实现[1]。智能和互联系统的承诺是，它们能够以更少的人力实现全新的功能。未来至关重要的系统--如自动驾驶汽车和能源网--将是智能和连接的。该提案的目标是大幅改善我们理解和设计这些智能连接系统的理论和工具。有两个主要的研究挑战使得构建这些系统变得困难。第一个是“解释”。解释是为系统用户和设计者准备的，它帮助我们理解软件是否实现了我们的意图。解释对于智能的、连接的系统至关重要，因为它们是可组合的（具有以前未知的、涌现的属性），而且它们很难解释，因为它们涉及复杂的机器学习算法。解释包括为什么系统采取它所做的动作，以及它在未来将如何改变。第二个主要的研究挑战是 * 设计 * 当前和现有的系统，以满足用户的意图。由于它们是可组合的和自主的，因此需要新的设计理论来约束这些系统如何发展。F-22战斗机80%的功能来自软件[9]。然而，将网络软件组件和Web服务接口（又称API）与机器学习领域的新兴工作相结合的结果有望大大加快软件主导我们世界的步伐。我的研究的影响将是更大的信心，在组装这些系统，并更好地了解如何设计它们，以支持用户的意图。智能和互联系统是先进的技术，需要知识和培训才能进行开发和操作。具有理解和构建这些系统的技能的工程师将至关重要。该提案将为智能互联系统开发新的设计方法。它将创建从用户意图到设计工件的跟踪工具。最后，该提案将研究人类如何解释这些系统的行为。

The characteristics of large, complex software systems are changing. Software is becoming more intelligent, adding machine learning components that are difficult for humans to interpret. Software is becoming more connected, formed of partially orchestrated sets of interacting components that create structures that are realized only at runtime [1]. The promise of intelligent and connected systems is that they enable radical new functionality with much less human effort. Future critically important systems---like autonomous vehicles and energy grids---will be intelligent and connected.The goal of this proposal is to dramatically improve the theories and tools we have for understanding and designing these intelligent, connected systems. There are two major research challenges that make constructing these systems difficult. The first is *explanation*. Explanation is for system users and designers, and helps us understand whether the software is fulfilling our intentions for it. Explanation is vital for intelligent, connected systems, since they are composable (with previously unknown, emergent properties), and they are hard to interpret, since they involve complex machine learning algorithms. Explanation includes why the system takes the actions it does, and how it will change in the future. The second major research challenge is *designing* current and existing systems to fulfil user intentions. Since they are composable and autonomous, new design theories will be needed to constrain how these systems evolve.The portion of a project's functionality (and cost) due to software is growing. 80% of functionality in the F-22 fighter jet is from software [9]. However, the consequences of combining networked software components and web service interfaces (aka APIs) with new and emerging work in machine learning promises to dramatically accelerate the pace with which software dominates our world. The impact of my research will be greater confidence in assembling these systems, and a better understanding of how to design them to support user intentions. Intelligent and connected systems are advanced technology that require knowledge and training for their development and operation. Engineers with the skills to understand and architect these systems will be vital. This proposal will develop new design approaches to intelligent, connected systems. It will create tools for tracing from user intentions to design artifacts. And finally, the proposal will investigate how humans explain the behaviour of these systems.