神经网络模拟框架柱低周反复加载试验性能初探

利用神经网络的智能特点，通过对一框架柱在低周反复水平荷载作用下行为的“学习”，来模拟框架柱经受不同荷载历史的性能。模拟结果试验检验，符合良好，表明了用神经网络模拟试验代替部分真实验是可行的。     

一种预测内燃机稳态排放特性的方法

为预测内燃机在额定工况范围内的排放特性，综合利用试验和模拟研究的优势，基于少量有代表性、易于实现的排放试验结果，利用神经网络的非线性映射能力，建立神经网络排放预测模型，研究结果表明该方法是可行的，提供了快速了解内燃机排放特性的手段。     

手写体数字识别的一种模糊神经网络方法

将模糊特征取技术与区组设计前馈网络相结合用于手写体数字识别。实验表明，该方法对５００个未学习样本的识别率达到了９５．８％。这一结果，相当于或优于目前已经发表的先进成果。     

基于MapInfo的城市公交查询系统的设计与实现

研究城市公交查询系统的设计与实现。利用动态分段技术,建立了基于MapInfo的动态分段数据结构,充分考虑乘客出行心理,提出了公交出行最优路径算法,设计了城市公交查询系统,实现输入查询信息或直接对地图操作来获得出行最优路线和换乘方案,提高乘客出行的便利性和高效性。     

一种改进的模型参考自适应算法

针对一种模型跟随自适应算法（DS-AMFC），利用线性神经网络预测系统输出与模型输出之差为控制信号的计算提供较准确的偏差值，用该估计的偏差量进行控制。同时，把控制的误差引入控制信号，在输入信号频率较高时仍具有跟随模型的能力，鲁棒性大大提高。     

关于神经计算中满意解的研究

在分析神经计算中已有的满意运算的基础上，找出其中存在的问题，提出了与阈值有关的整体满意度，并对其进行了分析比较，结果表明：整体满意度综合了组合满意度和平均满意度的优点，明显优于后两者。     

一个安全有效的会议密钥分配方案

提出了M．Steiner等人提出的会议密钥分配方案GDH．2存在的安全漏洞，并提出了一个新的安全有效的密钥分配方案，该方案适合多个用户通过不安全的通信网络进行信息交流。相对于GDH．2而言，本文案只以增加很小的计算量和通信负荷为代价，使安全性能得到较大的提高。     

Distance and time in intermodal goods transport networks in Europe: A generic approach

This paper is about distance and time as factors of competitiveness of intermodal transport. It reviews the relevance of the factors, evaluates time models in practice, compares network distances and times in alternative bundling networks with geometrically varied layouts, and points out how these networks perform in terms of vehicle scale, frequency and door-to-door time. The analysis focuses on intermodal transport in Europe, especially intermodal rail transport, but is in search for generic conclusions. The paper does not incorporate the distance and time results in cost models, and draws conclusions for transport innovation, wherever this is possible without cost modelling. For instance, the feature vehicle scale, an important factor of transport costs, is analysed and discussed.Distance and time are important factors of competitiveness of intermodal transport. They generate (direct) vehicle costs and – via transport quality – indirect costs to the customers. Clearly direct costs/prices are the most important performance of the intermodal transport system. The relevance of quality performances is less clarified. Customers emphasise the importance of a good match between the transport and the logistic system. In this framework (time) reliability is valued high. Often transport time, arrival and departure times, and frequency have a lower priority. But such conclusions can hardy be generalised. The range of valuations reflects the heterogeneity of situations. Some lack of clarity is obviously due to overlapping definitions of different performance types.The following parts of the paper are about two central fields of network design, which have a large impact on transport costs and quality, namely the design of vehicle roundtrips (and acceleration of transport speed) and the choice of bundling type: do vehicles provide direct services or run in what we call complex bundling networks? An example is the hub-and-spoke network. The objective of complex bundling is to increase vehicle scale and/or transport frequency even if network volumes are restricted. Complex bundling requires intermediate nodes for the exchange of load units. Examples of complex bundling networks are the hub-and-spoke network or the line network.Roundtrip and bundling design are interrelated policy fields: an acceleration of the roundtrip speed, often desirable from the cost point of view, can often only be carried out customer friendly, if the transport frequency is increased. But often the flow size is not sufficient for a higher frequency. Then a change of bundling model can be an outcome.Complex bundling networks are known to have longer average distances and times, the latter also due to the presence of additional intermediate exchange nodes. However, this disadvantage is – inside the limits of maximal vehicle sizes – overruled by the advantage of a restricted number of network links. Therefore generally, complex bundling networks have shorter total vehicle distances and times. This expression of economies of scale implies lower vehicle costs per load unit.The last part of the paper presents door-to-door times of load units of complex bundling networks and compares them with unimodal road transport. The times of complex bundling networks are larger than that of networks with direct connections, but nevertheless competitive with unimodal road transport, except for short distances.     

Leveraging social networks for efficient hurricane evacuation

One of the important factors affecting evacuation performance is the departure time choices made by evacuees. Simultaneous departures of evacuees can lead to overloading of road networks causing congestion. We are especially interested in cases when evacuees subject to little or no risk of exposure evacuate along with evacuees subject to higher risk of threat (also known as shadow evacuation). One of the reasons for correlated evacuee departures is higher perceived risk of threat spread through social contacts. In this work, we study an evacuation scenario consisting of a high risk region and a surrounding low risk area. We propose a probabilistic evacuee departure time model incorporating both evacuee individual characteristics and the underlying evacuee social network. We find that the performance of an evacuation process can be improved by forcing a small subset of evacuees (inhibitors) in the low risk area to delay their departure. The performance of an evacuation is measured by both average travel time of the population and total evacuation time of the high risk evacuees. We derive closed form expressions for average travel time for ER random network. A detailed experimental analysis of various inhibitor selection strategies and their effectiveness on different social network topologies and risk distribution is performed. Results indicate that significant improvement in evacuation performance can be achieved in scenarios where evacuee social networks have short average path lengths and topologically influential evacuees do not belong to the high risk regions. Additionally, communities with stronger ties improve evacuation performance.     

基于分段学习模型的自动驾驶行为决策算法研究

在具有车道线的特定自动驾驶场景中,针对目前端到端的行为决策算法直接输入原始图像进行决策导致的网络模型迁移性差、预测精度欠佳、泛化能力不足等问题,提出一种基于分段学习模型的车辆自动驾驶行为决策算法。首先,基于GoogLeNet建立一种端到端的车道线检测网络模型,并引入车道中心线作为决策的重要线索提高算法的迁移能力,同时利用YOLOv3目标检测模型对本车道内前方最近障碍物进行位置检测;而后,经几何测量模型将两者检测结果转换成环境状态信息向量为决策做支撑;最后,构建基于长短期记忆(LSTM)网络的驾驶行为决策模型,根据编码的历史状态信息刻画出动态环境中车辆的运动模式,并结合当前时刻的状态推理得到驾驶行为参量。使用建立的真实驾驶场景数据集对模型分别进行训练、验证与测试,离线测试结果显示车道线检测模型的检测位置误差小于1.3%,车道内前方障碍物检测模型的检测精度达98%以上,驾驶行为决策网络模型表征预测优度的决定系数 大于0.7。为进一步验证算法的有效性,搭建了Simulink/PreScan联合仿真平台,多种工况下的仿真验证试验中多个评价指标均达到工程精度要求,实车测试的试验结果也表明该算法可实现复杂驾驶场景下平稳、准确无偏航的预测效果并满足实时性要求,且与传统端到端模式的算法相比,具有更好的迁移性和泛化能力。