Goal-Driven Autonomy for Cognitive Systems
Goal-Driven Autonomy for Cognitive Systems
复制标题
认知系统的目标驱动自主性
DOI：
--
复制
发表时间：
2014
期刊：
影响因子：
--
通讯作者：
D. Perlis
中科院分区：
文献类型：
--
作者：
Matthew Paisner;Michael T. Cox;Michael Maynord;D. Perlis
认知系统的目标驱动的自主性Matt Paisner（mpaisner@umd.edu）Michael T.考克斯（mcox@cs.umd.edu）迈克尔·梅诺德（maynord@umd.edu）唐·佩里斯（perlis@cs.umd.edu）计算机科学系，A. V.威廉姆斯大楼。大学公园，MD 20742美国摘要复杂的，动态的环境提出了特殊的挑战，自主代理。具体来说，当世界不与设计假设合作时，代理人会遇到困难。我们提出了一种自治方法，旨在最大限度地提高鲁棒性，而不是在特定任务上的最优性。目标驱动的自主性包括识别可能的新问题，解释问题的原因并制定解决问题的目标。我们提出了这样一个模型中的MIDCA认知架构，并表明在某些条件下，该模型优于一个不太灵活的方法来处理意外事件。关键词：目标生成;异常处理;解释和解释; TF树;认知架构;智能自治。人类是非常多才多艺的，在处理各种意想不到的情况的同时，仍然在高层次的目标上取得进展。人类也可以在新的问题和机会出现时识别它们，并做出适当的反应。然而，在大多数情况下，我们的机器不能;它们就像白痴学者，擅长一项狭隘的任务，对其他任何事情都毫无用处，甚至与它们设计的任务非常相似。这就是所谓的脆性问题，这是AI的主要绊脚石。我们需要的似乎是与专家系统相反的东西：机器可能不擅长任何事情，但可以应付各种各样的情况，并保持战略眼光。然而，50多年的艰苦努力未能生产出这样的机器。解决脆弱性问题的一种方法是在认知架构中使用我们所说的目标驱动的自主性。在这里，我们描述了这种方法在动态环境中的一些好处。目标驱动的自治（GDA）是一个独特的概念，它赋予自治代理完全的独立性（考克斯，2007; Klenk，Molineaux，& Aha，2013; Munoz-Avila，Jaidee，Aha，Carter，2010）。代理不是任意的异常检测，而是在其当前目标和使命的上下文中搜索问题。并不是所有的异常都是问题，也不是所有的问题都重要到值得关注。代理人不应对整个世界状态进行一般性评估，而应溯因地解释引起问题的因果因素。给定解释，GDA代理可以生成解决问题的（可能是新的）目标（例如，通过移除其支持条件）。在这些方面，GDA是关于问题识别的，因为它是解决问题的（考克斯，2013）。考虑一下在建筑工地发生的火灾。这在许多方面都是一个问题，其中最重要的是行动的先决条件（例如，建筑材料的完整性）将变得不满意。因此，标准的规划算法可能会生成一个子目标来灭火，以便继续施工。然而，随后发生的火灾可能会证明对建筑工地的长期威胁进行调查是合理的。一个可能的解释是纵火犯的存在，导致犯罪者入狱的目标。因此，对火灾的直接反应方法是将其扑灭;对同样情况的GDA方法将认识到潜在问题对企业未来的威胁。本文将探讨这种方法之间的区别，智能推理和行为的元认知架构称为MIDCA，并将报告一个简单的实证研究的结果，以评估这些差异。在第2节中，我们介绍了MIDCA架构，其中包含GDA模型的实现实例。在第3节中，我们使用三种不同的目标生成方法来评估系统的性能：外生目标;统计生成的目标;由知识丰富的解释系统生成的目标。第4节概述了未来的工作，第5节调查相关的工作，第6节结束。认知架构中的目标驱动的自主性元认知、集成、双循环架构（MIDCA）（考克斯、梅诺德、佩斯纳、佩里斯和奥茨，2013年）由认知（即，对象）级和元语义（即，Meta）水平。图1显示了抽象了元级别的对象级别的已实现组件。每个循环的输出端包括意图、计划和行动执行，而输入端包括感知、解释和目标评估。一个循环选择一个目标并致力于实现它。然后，代理创建一个计划来实现目标，并随后执行计划的操作，使域匹配目标状态。智能体感知由行动引起的环境变化，解释关于计划的感知，并评估关于目标的解释。在
Goal-Driven Autonomy for Cognitive Systems Matt Paisner (mpaisner@umd.edu) Michael T. Cox (mcox@cs.umd.edu) Michael Maynord (maynord@umd.edu) Don Perlis (perlis@cs.umd.edu) Computer Science Department, A.V. Williams Bldg. College Park, MD 20742 USA Abstract Complex, dynamic environments present special challenges to autonomous agents. Specifically, agents have difficulty when the world does not cooperate with design assumptions. We present an approach to autonomy that seeks to maximize robustness rather than optimality on a specific task. Goal- driven autonomy involves recognizing possibly new problems, explaining what causes the problems and generating goals to solve the problems. We present such a model within the MIDCA cognitive architecture and show that under certain conditions this model outperforms a less flexible approach to handling unexpected events. Keywords: goal generation; anomaly handling; interpretation and explanation; TF-Tree; cognitive architecture; intelligent autonomy. Introduction Humans are astonishingly versatile, dealing with a wide range of unanticipated circumstances while still making headway on high-level goals. Humans can also recognize new problems and opportunities when they arise and react appropriately to them. Yet for the most part our machines cannot; they are like idiot-savants, very good at one narrow task and useless for anything else, even tasks very similar to the one they were designed for. This is the so-called brittleness problem, a major stumbling block for AI. What we appear to need is the opposite of expert systems: machines that might not excel at anything, but that can muddle through a wide range of circumstances and keep a strategic perspective. Yet more than 50 years of intense effort has failed to produce such machines. One approach to the problem of brittleness uses what we call goal-driven autonomy in a cognitive architecture. Here we describe some benefits of this approach in dynamic environments. Goal-Driven Autonomy (GDA) is a unique conception that gives full independence to autonomous agents (Cox, 2007; Klenk, Molineaux, & Aha, 2013; Munoz-Avila, Jaidee, Aha, Carter, 2010). Rather than arbitrary anomaly-detection, the agent searches for problems in the context of its current goals and mission. Not all anomalies are problems, nor are all problems important enough to attend to. Rather than general assessment of an entire world state, the agent should abductively explain the causal factors giving rise to the problem. Given an explanation, a GDA agent may generate a (possibly new) goal that solves the problem (e.g., by removing its supporting conditions). In these terms, GDA is as much about problem recognition as it is problem-solving (Cox, 2013). Consider a fire that breaks out at a construction site. This is a problem in many ways, not the least of which is that preconditions for actions (e.g., the integrity of building materials) will become unsatisfied. A standard planning algorithm might therefore generate a subgoal to extinguish the fire so that construction can continue. However a subsequent fire in quick succession might justify an investigation into a long-term threat to the construction site. One possible explanation would be the presence of an arsonist, leading to the goal of having the perpetrator in jail. Hence a direct reactive approach to fires would be to put them out; a GDA approach to this same situation would recognize the underlying problem in terms of its threat to the future of the enterprise. This paper will examine the distinction between such approaches to intelligent reasoning and behavior in a metacognitive architecture called MIDCA and will report the results of a simple empirical study to evaluate these differences. In section 2, we present the MIDCA architecture containing an implemented instantiation of the GDA model. In section 3 we evaluate the performance of systems making use of three distinct goal generation methods: exogenous goals; statistically generated goals; goals produced by a knowledge rich explanation system. Section 4 presents an overview of future work, section 5 surveys related work, and section 6 concludes. Goal-Driven Autonomy in a Cognitive Architecture The Metacognitive, Integrated, Dual-Cycle Architecture (MIDCA) (Cox, Maynord, Paisner, Perlis, & Oates, 2013) consists of “action-perception” cycles at both the cognitive (i.e., object) level and the metacognitive (i.e., meta-) level. Figure 1 shows the implemented components of the object level with the meta-level abstracted. The output side of each cycle consists of intention, planning, and action execution, whereas the input side consists of perception, interpretation, and goal evaluation. A cycle selects a goal and commits to achieving it. The agent then creates a plan to achieve the goal and subsequently executes the planned actions to make the domain match the goal state. The agent perceives changes to the environment resulting from the actions, interprets the percepts with respect to the plan, and evaluates the interpretation with respect to the goal. At the