CAREER: A Software Development Framework That Integrates Learning, Probabilistic Reasoning, And Any-Time Computation

越来越多的今天的计算机配备了传感器和执行器，这样它们就可以直接与人和外部世界互动。具体的例子包括通过互联网与人和其他软件代理交互的软件代理，以及与人和物理世界交互的消费产品。这类通常被称为“嵌入式”的计算机面临着各种各样的挑战。例如，传感器的限制可能会使其无法准确地感知环境的状态；因此，嵌入式计算机必须能够应对不确定性。嵌入式计算机的环境可能是动态的，在这种情况下，需要适应其中的变化。此外，嵌入式计算机可能必须满足实时限制，因为它们的环境往往不会等待其程序的终止。虽然其中许多问题已经在人工智能(AI)领域进行了研究，但计算机科学仍然缺乏一种完善的嵌入式计算机软件开发方法。现有的编程语言和工具包没有解决不确定性、适应性和实时计算等问题。PI的职业目标是改变嵌入式计算机的编程方式。为了实现这一目标，他提议研究一种专门为嵌入式计算机系统设计的新编程语言。这种语言将引入三个目前在现有编程语言中找不到的新概念：I.概率数据类型和运算符。概率数据类型和运算符将使程序员能够使用不确定的信息进行计算，就像它是确定的一样。概率数据类型通过概率密度表示信息。它们概括了现有的数据类型(例如，FLOAT、INT)，因为它们一次表示多个值，由数字概率加权。从程序员的观点来看，概率计算将类似于传统计算，并增加了对不确定性的健壮性。适应。PI将开发支持程序代码的数据驱动适配的机制。语言将拥有内置的学习算法(例如，神经网络、强化学习)和适当的学分分配机制。这将支持可适应软件的设计，能够自动适应环境的变化。它还将使程序员能够“教授”他们的代码，作为对当前软件开发实践的补充。随时行刑。为了适应及时响应的需要，新语言将为任何时间计算提供运行时支持。可以(几乎)在任何时间查询任何时间程序的结果；然而，它们的解决方案的质量会随着时间的推移而提高。为了获得任意时间特征，执行系统将按单个值的概率降序选择性地处理概率变量。将提供一种特殊的机制，允许以事件驱动的方式终止任何时间的计算。如果这项研究成功，嵌入式系统的程序员将能够以最小的开销利用结果，与当今的最佳实践相比，有效地使他们能够以更少的努力开发更好的软件。为了更好地实现他的职业目标，PI还将参与一系列教育活动。他计划修改CMU编程入门课程的课程，增加一个关于嵌入式计算的课程。他还将开发一门新的研究生级别的嵌入式计算课程，该课程将整合目前不同学科的材料。在这两门课程中，编程语言将被用作工具：通过使用这种语言，学生将对嵌入式计算涉及的各种问题及其理论基础有更深入的了解。此外，高级学生的反馈将指导这里提出的基础研究。所有的课程材料、语言、实施和文件以及所有相关的研究论文都将通过网络向广大研究界提供，以便其他人能够为该项目作出贡献并从中受益。

An increasing number of today's computers are equipped with sensors and actuators, so that they can interact directly with people and the outside world. Specific examples include software agents that interact with people and other software agents through the Internet, and consumer products, which interact with people and the physical world. Such computers, often called "embedded," face a variety of challenges. For example, sensor limitations might make it impossible to sense the state of the environment accurately; thus, embedded computers must be able to cope with uncertainty. The environments of embedded computers might be dynamic, in which case there is a need to adapt to changes therein. Additionally, embedded computers might have to meet real-time constraints, as their environments tend not to wait for the termination of their programs. While many of these issues have been researched in the field of artificial intelligence (AI), computer science still lacks a sound methodology for software development in embedded computers. Existing programming languages and toolkits do not address issues such as uncertainty, adaptation, and on-time computing.The PI's career objective it to change the way embedded computers are programmed. To pursue this goal, he proposes research towards a new programming language, specifically designed for embedded computer systems. This language will introduce three new ideas currently not found in existing programming languages:I. Probabilistic data types and operators. Probabilistic data types and operators will enable programmers to compute with uncertain information just as if it was certain. Probabilistic data types represent information by probability densities. They generalize existing data types (e.g., float, int) in that they represent multiple values at a time, weighted by numerical probabilities. From the programmer's view, probabilistic computation will be analogous to conventional computation, with the added benefit of increased robustness to uncertainty.2. Adaptation. The PI will develop mechanisms that support data-driven adaptation of program code. language will possess built-in learning algorithms (e.g., neural networks, reinforcement learning) and a mechanism for proper credit assignment. This will support the design of adaptable software, capable of automatically adapting to changes in the environment. It will also enable programmers to "teach" their code, as a supplement to current software development practice.3. Any-time execution. To accommodate the need for timely responses, the new language will provide run-time support for any-time computation. Any-time programs can be queried for a result at (almost) any time; the quality of their solutions, however, increases over time. To obtain any-time characteristics, the execution system will process probabilistic variables selectively, in decreasing order of the probability of individual values. A special mechanism will be provided that permits the event-driven termination of any-time computation.If this research is successful, programmers of embedded systems will be able to utilize the results-with minimal overhead, effectively enabling them to develop better software with significantly less effort, when compared to today's best practice.To better achieve his career goal, the PI will also engage in a collection of educational activities. He plans to revise the curriculum of CMU's introductory programming class, by including a course segment on embedded computation. He will also develop a new graduate-level course on embedded computation, which will integrate material from a variety of currently separate disciplines. For both courses, the programming language will be used as a vehicle: By using this language, students will gain a much deeper understanding of the various issues involved in embedded computation and their theoretical foundations. Additionally, feedback of advanced students will guide the basic research proposed here. All course materials, the language, its implementation and documentation, and all related research papers will be made available to the research community at large using the Web, so that others can contribute to and benefit from this project.