Multiagent Search Algorithms for Learning & Planning in Colony-Style Robots

在智能机器人系统中直接实现感觉-认知运动三位一体通常会对这些系统的计算、通信和存储资源提出很高的要求。这个问题的一个解决方案是在机器人的不同物理部件之间分配三合一的功能。群居式机器人ı1！和它的前身ı2！(小计体系结构)代表了实现这种功能分布的相关方法。群集式机器人使用一组(即，群集式)分级禁止的代理来控制机器人。现有的群集式机器人存在一些缺点，限制了它们作为自主系统的应用。这些缺点涉及确定机器人的基本控制等级的问题，以及现有系统无法保存和使用先前的经验。这个项目使用了多代理处理范例，称为多代理搜索策略的大规模算法ı3！解决现行群体式体制的不足。大量算法的灵感来自于最近关于竞争和合作抑制神经网络ı3！ı4！的工作。利用统计力学论证，可以证明竞争质量算法可以形成密度函数的最小熵表示。该项目使用分析形式主义来开发能够学习群体式控制层次结构的组件和结构的大量算法。研究还探讨了如何利用“协作”质量来实现Bellman动态规划方程的完全离散化求解，并利用这种能力设计了一种在群集式机器人中实现长期记忆和路径规划能力的方法。该项目将通过模拟和分析，充分开发在群集式机器人中实现记忆、学习和规划所需的大量算法。这项研究还将研究将这种方法扩展到更复杂的系统，如柔性制造系统。

Direct implementation of the sensory-cognitive motor triad for intelligent robotic systems will generally place severe demands on the computing, communication, and memory resources of these systems. One solution to this problem has been to distribute the triad's functionality among various physical components of the robot. The colony-style robot ı1! and its predecessor ı2! (subsummation architecture) represent related methods for achieving this functional distribution. Colony-style robots use a collection(ie., colony) of hierarchically inhibited agents to control the robot. Existing colony-style robots suffer several disadvantages limiting their utility as autonomous system. These disadvantages involve problems in determining the robot's underlying control hierarchy and the inability of existing systems to save and use prior experience. This project uses multiagent processing paradigms called multiagent search strategies of MASS algorithms ı3! to solve the deficiencies of current colony-style systems. MASS algorithms ar inspired by recent work with competitively and cooperatively inhibited neural networks ı3!ı4!. It can be shown, using statistical mechanical arguments, that competitive MASS algorithms can form minimum entropy representations of density functions. This project uses that analytical formalism to develop MASS algorithms capable of learning the components and structure of the colony-style control hierarchy. The research also explores how "cooperative" MASS can be used to realize fully discretized solution of Bellman's dynamic programming equation, and uses this capability to devise a method for implementing long term memory and path planning capabilities in colony-style robots. The project will fully develop, through simulation and analysis, the MASS algorithms needed to realize memory, learning, and planning in colony-style robots. The research will also investigate extensions of this approach to more complex systems such as flexible manufacturing systems.