Learning and Intelligent Systems: Agile Procedural Learning Systems

对学习的研究表明，人们可以定义两大类：陈述性学习(例如，口头陈述事实)和程序性学习(例如，如何骑自行车)。本研究旨在阐明程序性学习的原理，从而对理解如何训练人或机器在变化的环境中完成复杂的感觉运动任务产生影响。研究人员的目标是将对动物和机器人的研究结合起来，以探索一种以生物为灵感的敏捷程序学习系统(APLS)的架构。现有的敏捷程序性学习系统包括人类和动物。即使是更简单的动物，如昆虫或软体动物，在解决生存问题方面也表现出非凡的敏捷性。在遇到新的地形或新的食物后，这些动物可以调整身体或进食运动的模式，以迅速应对新的情况。将受益于敏捷程序学习理论的工程系统示例包括机器人和制造机床。特别是考虑到敏捷制造的应用领域。敏捷制造机床的关键要求是能够灵活、快速地适应不断变化的装配线需求、感知到的错误和新的任务。尽管机械手变得更灵活，传感器变得更准确，尽管处理器变得更快，内存更便宜，软件更容易编写和维护，但学习和展示智能行为的真正敏捷的工程系统还没有得到证明。底物并不缺乏；需要一种能够产生工程原理的理论方法。学习和智能系统(LIS)的自上而下(认知)和自下而上(反应)方法已经取得了成功，但分别适用于在高度结构化环境中执行高级任务的高度复杂的程序和在温和变化的环境中执行低级任务的简单代理。相比之下，敏捷的程序性学习系统在反应行为和认知技能之间具有中等水平的能力。它们是复杂的、受限的，并且必须在结构变化的环境中工作。因此，敏捷过程性学习的问题直接摆在研究人员面前，他们需要建立一种LIS理论来弥合传统的“自上而下”和“自下而上”方法之间的差距。研究人员提出的缺失的“中间出”理论将采用以下方法：(1)确定现有敏捷程序团队系统中使用的原则；(2)开发模型和数学工具，弥合“自上而下”和“自下而上”方法之间的差距；以及(3)将这些发现转化为工程实践。更具体地说，他们将探索以下受生物启发的体系结构，作为解决灵活程序性学习问题的一种手段：塑料局部反射电路，由更高级别的状态和环境相关电路协调和调制。研究人员的研究工作有三个核心推动力：(1)他们将在能够进行程序性学习的更简单的动物身上进行实验研究，以确定这种体系结构实际上是如何用于学习任务的；(2)他们将开发模型和数学工具，将这些发现形式化，使综合的符号反应理论成为可能，并提高对学习和环境对智能自主代理动态影响的理解；以及(3)他们将把中间输出理论和体系结构纳入类似动物的机器人和敏捷制造车间，通过提高性能来评估它们在产生学习和智能行为方面的效用。中间输出方法产生了关于低级感觉运动学习和高级学习的主要科学问题，这些问题只能用跨学科的方法来回答。将这些努力相互联系起来的重叠的任务小组将协同改进每一项工作。

Studies of learning indicate that one can define two broad categories: declarative learning (e.g., verbal statements of fact) and procedural learning (e.g., how to ride a bicycle). This research aims to clarify principles of procedural learning, and thus have an impact on understanding how to train people or machines to accomplish complex sensorimotor tasks in changing environments. The researcher's goal is to combine studies on animals and robots in order to explore a biologically-inspired architecture for agile procedural learning systems (APLS). Existing agile procedural learning systems include humans and animals. Even simpler animals, such as insects or mollusks, show a remarkable agility in solving survival problems. After encountering novel terrain, or novel food, these animals can adjust patterns of body or ingestive movements to handle the new situation rapidly. Example engineering systems that would benefit from a theory of agile procedural learning include robots and manufacturing workcells. In particular, considering the application domain of agile manufacturing. The key requirement of an agile manufacturing workcell is that it be capable of flexibly and rapidly adjusting to changing assembly line demands, perceived errors, and new tasks. Though manipulators have gotten more dextrous and sensors more accurate, though processors have gotten faster, memory cheaper, and software easier to write and maintain, truly agile engineering systems that learn and exhibit intelligent behavior have not been demonstrated. The substrate is not lacking; a theoretical approach yielding engineering principles is needed. Top-down (cognitive) and bottom-up (reactive) approaches to learning and intelligent systems (LIS) have yielded successes, but only for highly-complex programs performing high-level tasks in highly-structured environments and for simple agents performing low-level tasks in mildly-changing environments, respectively. In contrast, agile procedural learning systems possess a middle-level competence between reactive behavior and cognitive skills. They are complex, constrained, and must work in environments whose structure changes. Thus, the problem of agile procedural learning confronts the investigators directly with building a theory of LIS that bridges the gap between the traditional "top down" and "bottom up" approaches. The missing "middle out" theory the investigators propose will be pursued with the following approach: (1) identify principles used in existing agile procedural teaming systems; (2) develop models and mathematical tools that bridge the gap between "top-down" and "bottom-up" methods; and (3) transfer these findings to engineering practice. More specifically, they will explore the following biologically-inspired architecture as a means of solving the problem of agile procedural learning: plastic local reflex circuitry coordinated and comodulated by higher-level state- and environment-dependent circuits. The investigators research effort has three core thrusts: (1) they will pursue experimental studies in simpler animals that are capable of procedural learning to determine how this architecture is actually used for learning tasks; (2) they will develop models and mathematical tools that formalize these findings, make an integrative symbolic-reactive theory possible, and improve understanding of the influence that learning and environment have on the dynamics of intelligent autonomous agents; and (3) they will incorporate both into animal-like robots and an agile manufacturing workcell the middle-out theory and architecture, evaluating their utility in producing learning and intelligent behavior, measured by improved performance. The middle-out approach generates overarching scientific questions about low-level sensorimotor learning and higher-order learning that can only be answered using an interdisciplinary approach. Overlapping task teams interlinking these efforts will synergistically improve each of them.