CAREER: Neural Network-Based Adaptive Critic Designs for Broadband Network Traffic Control

9874601 liu自适应批评设计（ACDs）是一般情况下近似动态规划的设计。典型的ACD由三个可以使用神经网络实现的模块组成——值模块、预测模块和决策模块。本计划旨在与业界合作，利用ACD为宽频通讯网络开发和实施学习流量控制方案。下一代通信网络设计者面临的一个开放性和挑战性的问题是如何设计有效地集成多媒体流量并保证每个流量源的服务质量（QoS）的方案。这里的挑战是最大限度地利用网络带宽，同时保证QoS。为此，包括呼叫接纳控制和交通执法在内的交通控制方案在过去得到了广泛的研究。然而，现有的大多数方案为了保证QoS而牺牲带宽利用率（即浪费网络资源/损失收益）。该项目的初步研究表明，在相同的QoS目标下，使用基于ACDs的自学习方法可以实现比现有方案更高的带宽利用率。本项目将实现不同级别（分组/呼叫/网络级别）的综合自学习流量控制。其中包括具有优化参数的令牌桶、呼叫准入控制以及路由和拥塞控制。本文将研究基于神经网络的ACDs领域的两个基本问题：(1)基于神经网络的ACDs的结构鲁棒性分析；(2)神经网络训练中的收敛问题。教育活动包括：(1)课程、实验室和课程开发：开发一门关于ACDs的研究生课程，开发史蒂文斯智能系统实验室，以及跨学科的努力，将重点放在ACDs/类脑智能系统上；(2)本科生的研究经验：让本科生接触到各种ACDs研究领域的众多机会；(3)教学措施：开发基于网络的交互式计算机工具，用于自动提交作业和评分，并在计算机网络上实施白板系统，以帮助学生学习；对教学措施的评估将采用基于网络的在线教育评估系统，该系统将ACDs的自我学习特性应用于最佳决策。***

9874601LiuAdaptive critic designs (ACDs) are designs that approximate dynamic programming in the general case. A typical ACD consists of three modules that can be implemented using neural networks -- a Values module, a Prediction module , and a Decision module. This project aims to collaborate with industry in developing and implementing learning traffic control schemes for broadband communication networks using ACD's.An open and challenging problem facing the designer of the next generation of communication networks is to design schemes that integrate multimedia traffic efficiently and that guarantee the quality of service (QoS) for each traffic source. The challenge here is to maximize the network bandwidth utilization and at the same time to guarantee QoS. For this purpose, traffic control schemes including call admission control and traffic enforcement have been studied extensively in the past. However, most of the existing schemes sacrifice bandwidth utilization (i.e., waste network resources/lose revenue) in order to guarantee QoS. The preliminary study of this project has shown that, with the same QoS goals, higher bandwidth utilization than existing schemes can be achieved using a self-learning approach based on ACDs.Integrated self-learning traffic control at different levels (packet/call/network levels) will be implemented in this project. These include token bucket with optimized parameters, call admission control, and routing and congestion control. Two fundamental issues in the field of neural network-based ACDs will be investigated: (1) structural robustness analysis of neural network-based ACDs and (2) problems relating to convergence in neural network training. Education activities include: (1) Course, laboratory and curriculum development: To develop a graduate course on ACDs, to develop the Intelligent Systems Laboratory at Stevens, and the interdisciplinary effort which will emphasize ACDs/brain-like intelligent systems; (2) Research experience for undergraduate students: To expose undergraduate students to the numerous Opportunities available in a variety of ACDs research areas; and (3) Pedagogical initiatives: To develop a web-based interactive computer tool for automated homework submission and grading and to implement a white board system over the computer network to aid student learning; the evaluation of the pedagogical initiatives will use a web-based, on-line education assessment system in which the self-learning feature of ACDs is applied to optimal decision making.***