Reinforcement-Learning Connectionist Systems (Computer and Information Science)

由类神经元处理单元组成的网络，称为连接主义系统，具有有趣的计算特性，使其在心理建模和人工智能的潜在应用方面都具有吸引力。最近的研究已经导致了一些在这样的网络中学习的有前途的算法。这个项目通过为随机单元的连接式网络中的强化学习这一特定问题开发一种具有良好数学基础的算法设计方法来推动这方面的研究。开发了用于学习具有重要时间分量的问题的特定算法，例如具有未知持续时间的反馈延迟的控制问题或涉及时变信号的识别或产生的问题。这些算法需要允许适当的在线实现，其中学习发生在操作系统中。除了在这类算法的理论上取得进展外，该项目还包括实现更有前途的候选算法并在模拟实验中评估它们的性能。评估的主要标准是学习效率和是否有可能收敛到次优状态。此外，该项目还通过对人工智能和机器人中特定问题的适当缩小版本的研究，探索了这些算法对这些问题的适用性。这些问题包括语音识别和自适应感觉运动控制。

Networks of neuron-like processing units, called connectionist systems, have interesting computational properties making them attractive for both psychological modeling and potential application in artificial intelligence. Recent research has led to some promising algorithms for learning in such networks. This project advances such research by developing a mathematically well-founded approach to the design of algorithms for the particular problem of reinforcement learning in connectionist networks of stochastic units. Particular algorithms are developed for learning problems having an important temporal component, such as control problems with feedback delays of unknown duration or problems involving recognition or production of time-varying signals. These algorithms are required to admit a suitable on-line implementation, in which the learning occurs within the operating system. In addition to making advances in the theory of such algorithms, this project involves implementing the more promising candidates and evaluating their performance in simulation experiments. Primary criteria for evaluation are learning efficiency and suspectibility to convergence to suboptimal states. In addition, this project explores the applicability of such algorithms to specific problems in artificial intelligence and robotics through the study of suitably scaled-down versions of these problems. Such problems include speech recognition and adaptive sensorimotor control.