随机动态规划的混合算法研究

针对随机条件下动态规划模型的主要特点,运用智能算法混合编程理论,设计了一种探索多阶段决策问题的智能混合算法.该算法首先将问题转化成一族同类型的一步决策子问题,然后利用随机模拟和遗传算法,依据训练样本形成的训练神经元网络,在单步决策中寻求最优策略和最优目标值,逐个求解,再据初始状态逆序求出最优策略序列和最优目标值.仿真结果表明,该算法具有一定的通用性,初始设计点可以随机产生,其计算精度不因函数的非线性强弱而受影响,对目标和约束的限制较少,可应用于多种形式的随机多阶段决策优化问题,较好地满足了随机动态规划模型求解和优化的要求.     

Stochastic regret minimization for revenue management problems with nonstationary demands

We study an admission control model in revenue management with nonstationary and correlated demands over a finite discrete time horizon. The arrival probabilities are updated by current available information, that is, past customer arrivals and some other exogenous information. We develop a regret‐based framework, which measures the difference in revenue between a clairvoyant optimal policy that has access to all realizations of randomness a priori and a given feasible policy which does not have access to this future information. This regret minimization framework better spells out the trade‐offs of each accept/reject decision. We proceed using the lens of approximation algorithms to devise a conceptually simple regret‐parity policy. We show the proposed policy achieves 2‐approximation of the optimal policy in terms of total regret for a two‐class problem, and then extend our results to a multiclass problem with a fairness constraint. Our goal in this article is to make progress toward understanding the marriage between stochastic regret minimization and approximation algorithms in the realm of revenue management and dynamic resource allocation. © 2016 Wiley Periodicals, Inc. Naval Research Logistics 63: 433–448, 2016     

A worst‐case formulation for constrained ranking and selection with input uncertainty

In this research, we consider robust simulation optimization with stochastic constraints. In particular, we focus on the ranking and selection problem in which the computing time is sufficient to evaluate all the designs (solutions) under consideration. Given a fixed simulation budget, we aim at maximizing the probability of correct selection (PCS) for the best feasible design, where the objective and constraint measures are assessed by their worst‐case performances. To simplify the complexity of PCS, we develop an approximated probability measure and derive the asymptotic optimality condition (optimality condition as the simulation budget goes to infinity) of the resulting problem. A sequential selection procedure is then designed within the optimal computing budget allocation framework. The high efficiency of the proposed procedure is tested via a number of numerical examples. In addition, we provide some useful insights into the efficiency of a budget allocation procedure.     

Maneuvering target tracking of UAV based on MN-DDPG and transfer learning

Tracking maneuvering target in real time autonomously and accurately in an uncertain environment is one of the challenging missions for unmanned aerial vehicles(UAVs).In this paper,aiming to address the control problem of maneuvering target tracking and obstacle avoidance,an online path planning approach for UAV is developed based on deep reinforcement learning.Through end-to-end learning powered by neural networks,the proposed approach can achieve the perception of the environment and continuous motion output control.This proposed approach includes:(1)A deep deterministic policy gradient(DDPG)-based control framework to provide learning and autonomous decision-making capa-bility for UAVs;(2)An improved method named MN-DDPG for introducing a type of mixed noises to assist UAV with exploring stochastic strategies for online optimal planning;and(3)An algorithm of task-decomposition and pre-training for efficient transfer learning to improve the generalization capability of UAV's control model built based on MN-DDPG.The experimental simulation results have verified that the proposed approach can achieve good self-adaptive adjustment of UAV's flight attitude in the tasks of maneuvering target tracking with a significant improvement in generalization capability and training efficiency of UAV tracking controller in uncertain environments.     

具有战时随机延误与损耗的多配送中心路径优化

建立了具有战时随机延误与损耗的多配送中心配送路径安排模型,给出了基于随机模拟的蚁群算法。算法通过给定残存率、用时与置信度阈值,把多目标问题作为单目标来处理。用随机模拟的方法来求路径的置信度,并以此为基础搜索转移策略的临域与判断未遍历点的插入位置。算法设计了符合问题特点的从虚拟点出发的转移策略与对两类路段不同的信息素更新策略,确保算法的实现。最后,通过算例说明了该方法的可行性与有效性。     

Dangerous work: Terrorism against U.S. diplomats

The goal of diplomats is to represent their countries’ interests through diplomacy, not arms. Because they are not military personnel, they may be perceived as at lower risk of being the target of terrorists. However, recent events have called this perception into question. Despite this danger, there has been little research on terrorist attacks against diplomats. Drawing on the terrorism studies literature, this article argues that diplomats are targeted more than non-diplomatic targets in countries where certain U.S. foreign policies are implemented. An empirical analysis of 471 attacks against U.S. diplomats from 1970 to 2011 reveals that while U.S. alliances and foreign aid increase the likelihood of attacks against diplomats, U.S. military intervention and civil war, on the other hand, increase the risk of terrorism against non-diplomatic targets. This finding is relevant because it shows terrorist attacks against diplomats result from certain types of foreign policy.     

Approximation algorithms for sequential batch‐testing of series systems

We introduce and study a generalization of the classic sequential testing problem, asking to identify the correct state of a given series system that consists of independent stochastic components. In this setting, costly tests are required to examine the state of individual components, which are sequentially tested until the correct system state can be uniquely identified. The goal is to propose a policy that minimizes the expected testing cost, given a‐priori probabilistic information on the stochastic nature of each individual component. Unlike the classic setting, where variables are tested one after the other, we allow multiple tests to be conducted simultaneously, at the expense of incurring an additional set‐up cost. The main contribution of this article consists in showing that the batch testing problem can be approximated in polynomial time within factor , for any fixed . In addition, we explain how, in spite of its highly nonlinear objective function, the batch testing problem can be formulated as an approximate integer program of polynomial size, while blowing up its expected cost by a factor of at most . Finally, we conduct extensive computational experiments, to demonstrate the practical effectiveness of these algorithms as well as to evaluate their limitations. © 2016 Wiley Periodicals, Inc. Naval Research Logistics 63: 275–286, 2016     

Adaptive importance sampling for extreme quantile estimation with stochastic black box computer models

Quantile is an important quantity in reliability analysis, as it is related to the resistance level for defining failure events. This study develops a computationally efficient sampling method for estimating extreme quantiles using stochastic black box computer models. Importance sampling has been widely employed as a powerful variance reduction technique to reduce estimation uncertainty and improve computational efficiency in many reliability studies. However, when applied to quantile estimation, importance sampling faces challenges, because a good choice of the importance sampling density relies on information about the unknown quantile. We propose an adaptive method that refines the importance sampling density parameter toward the unknown target quantile value along the iterations. The proposed adaptive scheme allows us to use the simulation outcomes obtained in previous iterations for steering the simulation process to focus on important input areas. We prove some convergence properties of the proposed method and show that our approach can achieve variance reduction over crude Monte Carlo sampling. We demonstrate its estimation efficiency through numerical examples and wind turbine case study.     

A sequential partial information bomber-defender shooting problem

A bomber carrying homogenous weapons sequentially engages ground targets capable of retaliation. Upon reaching a target, the bomber may fire a weapon at it. If the target survives the direct fire, it can either return fire or choose to hold fire (play dead). If the former occurs, the bomber is immediately made aware that the target is alive. If no return fire is seen, the true status of the target is unknown to the bomber. After the current engagement, the bomber, if still alive, can either re-engage the same target or move on to the next target in the sequence. The bomber seeks to maximize the expected cumulative damage it can inflict on the targets. We solve the perfect and partial information problems, where a target always fires back and sometimes fires back respectively using stochastic dynamic programming. The perfect information scenario yields an appealing threshold based bombing policy. Indeed, the marginal future reward is the threshold at which the control policy switches and furthermore, the threshold is monotonic decreasing with the number of weapons left with the bomber and monotonic nondecreasing with the number of targets left in the mission. For the partial information scenario, we show via a counterexample that the marginal future reward is not the threshold at which the control switches. In light of the negative result, we provide an appealing threshold based heuristic instead. Finally, we address the partial information game, where the target can choose to fire back and establish the Nash equilibrium strategies for a representative two target scenario.     

The (s,Q) inventory model with Erlang lead time and deterministic demand

We develop a simple, approximately optimal solution to a model with Erlang lead time and deterministic demand. The method is robust to misspecification of the lead time and has good accuracy. We compare our approximate solution to the optimal for the case where we have prior information on the lead‐time distribution, and another where we have no information, except for computer‐generated sample data. It turns out that our solution is as easy as the EOQ's, with an accuracy rate of 99.41% when prior information on the lead‐time distribution is available and 97.54–99.09% when only computer‐generated sample information is available. Apart from supplying the inventory practitioner with an easy heuristic, we gain insights into the efficacy of stochastic lead time models and how these could be used to find the cost and a near‐optimal policy for the general model, where both demand rate and lead time are stochastic. © 2004 Wiley Periodicals, Inc. Naval Research Logistics, 2004     

Using stochastic programming to solve an outpatient appointment scheduling problem with random service and arrival times

We study a stochastic outpatient appointment scheduling problem (SOASP) in which we need to design a schedule and an adaptive rescheduling (i.e., resequencing or declining) policy for a set of patients. Each patient has a known type and associated probability distributions of random service duration and random arrival time. Finding a provably optimal solution to this problem requires solving a multistage stochastic mixed‐integer program (MSMIP) with a schedule optimization problem solved at each stage, determining the optimal rescheduling policy over the various random service durations and arrival times. In recognition that this MSMIP is intractable, we first consider a two‐stage model (TSM) that relaxes the nonanticipativity constraints of MSMIP and so yields a lower bound. Second, we derive a set of valid inequalities to strengthen and improve the solvability of the TSM formulation. Third, we obtain an upper bound for the MSMIP by solving the TSM under the feasible (and easily implementable) appointment order (AO) policy, which requires that patients are served in the order of their scheduled appointments, independent of their actual arrival times. Fourth, we propose a Monte Carlo approach to evaluate the relative gap between the MSMIP upper and lower bounds. Finally, in a series of numerical experiments, we show that these two bounds are very close in a wide range of SOASP instances, demonstrating the near‐optimality of the AO policy. We also identify parameter settings that result in a large gap in between these two bounds. Accordingly, we propose an alternative policy based on neighbor‐swapping. We demonstrate that this alternative policy leads to a much tighter upper bound and significantly shrinks the gap.     

基于Fisher线性判别模型的文本特征选择算法

在采用向量空间模型表示方法的文本分类系统中,维数约简是必要的步骤,特征选择方法由于计算复杂度较低而被广泛采用.本文基于Fisher线性判别模型提出了一种新的文本特征选择算法,将其求解过程转换为一个特征项优化组合的问题,避免了复杂的矩阵变换运算.实验表明,该方法与信息增益、卡方统计方法比较,具有较明显的优势.     

Contractual agreements for coordination and vendor‐managed delivery under explicit transportation considerations

We consider the coordination problem between a vendor and a buyer operating under generalized replenishment costs that include fixed costs as well as stepwise freight costs. We study the stochastic demand, single‐period setting where the buyer must decide on the order quantity to satisfy random demand for a single item with a short product life cycle. The full order for the cycle is placed before the cycle begins and no additional orders are accepted by the vendor. Due to the nonrecurring nature of the problem, the vendor's replenishment quantity is determined by the buyer's order quantity. Consequently, by using an appropriate pricing schedule to influence the buyer's ordering behavior, there is an opportunity for the vendor to achieve substantial savings from transportation expenses, which are represented in the generalized replenishment cost function. For the problem of interest, we prove that the vendor's expected profit is not increasing in buyer's order quantity. Therefore, unlike the earlier work in the area, it is not necessarily profitable for the vendor to encourage larger order quantities. Using this nontraditional result, we demonstrate that the concept of economies of scale may or may not work by identifying the cases where the vendor can increase his/her profits either by increasing or decreasing the buyer's order quantity. We prove useful properties of the expected profit functions in the centralized and decentralized models of the problem, and we utilize these properties to develop alternative incentive schemes for win–win solutions. Our analysis allows us to quantify the value of coordination and, hence, to identify additional opportunities for the vendor to improve his/her profits by potentially turning a nonprofitable transaction into a profitable one through the use of an appropriate tariff schedule or a vendor‐managed delivery contract. We demonstrate that financial gain associated with these opportunities is truly tangible under a vendor‐managed delivery arrangement that potentially improves the centralized solution. Although we take the viewpoint of supply chain coordination and our goal is to provide insights about the effect of transportation considerations on the channel coordination objective and contractual agreements, the paper also contributes to the literature by analyzing and developing efficient approaches for solving the centralized problem with stepwise freight costs in the single‐period setting. © 2006 Wiley Periodicals, Inc. Naval Research Logistics, 2006     

Optimal selection with alternative information

This article examines the problem of optimally selecting from several unknown rewards when there are given alternative, costly sources of information. The optimal rule, indicating the information to be purchased and the reward to be selected, is specified as a function of the decision maker's prior probabilities regarding the value of each alternative. The rule is surprisingly complex, balancing prior beliefs, the “informativeness” of the relevant information system, and the cost of acquiring information.     

结合规则知识与神经网络于空袭武器的辅助选择

结合规则推理与神经网络,提出了一种新的空袭武器辅助选择方法。其根据空袭武器选择的基本原则以及问题的特点,构造T-S模糊逻辑系统及其实现的相应神经网络模型,然后通过已有的战例与试验例子训练神经网络,并对问题进行求解,得出空袭武器选择提供依据(一种量化的武器适用度)。另外,还利用专家提供的规则知识于学习实例的预处理,学习实例信息的扩充。     

Evaluating appointment postponement in scheduling patients at a diagnostic clinic

Diagnostic clinics are among healthcare facilities that suffer from long waiting times which can worsen medical outcomes and increase patient no-shows. Reducing waiting times without significant capital investments is a challenging task. We tackle this challenge by proposing a new appointment scheduling policy that does not require significant investments for diagnostic clinics. The clinic in our study serves outpatients, inpatients, and emergency patients. Emergency patients must be seen on arrival, and inpatients must be given next day appointments. Outpatients, however, can be given later appointments. The proposed policy takes advantage of this by allowing the postponement of the acceptance of appointment requests from outpatients. The appointment scheduling process is modeled as a two-stage stochastic programming problem where a portion of the clinic capacity is allocated to inpatients and emergency patients in the first stage. In the second stage, outpatients are scheduled based on their priority classes. After a detailed analysis of the solutions obtained from the two-stage stochastic model, we develop a simple, non-anticipative policy for patient scheduling. We evaluate the performance of this proposed, easy-to-implement policy in a simulation study which shows significant improvements in outpatient indirect waiting times.     

EUROPEAN DEFENCE FIRMS: THE INFORMATION BARRIER ON PRIVATE FINANCE*

After the fall of the Berlin Wall, European governments adopted a hands‐off policy towards the defence industrial base, in an attempt to increase the sector’s efficiency and reactivity. In this context, one topical issue is how to motivate defence firms to apply for private rather than public finance. Since banks have no prior experience with European defence firms, a problem of asymmetric information may block this transition. The problem is analysed within the framework of a game between defence firms and banks. It is shown that the Bayesian Equilibrium might correspond to a situation where low‐risk firms prefer the state‐financed scheme; yet, in a perfect information set‐up, the same firms would apply for bank credit. In order to facilitate the transition to private finance, the government might decide to subsidize investors who agree on financing defence firms; the state aid should be made available during a transitory learning period.     

基于迭代学习观测器的卫星姿态滑模容错控制

为解决卫星上反作用飞轮存在安装偏差、故障及外部干扰情况下的姿态控制问题,提出了一种基于迭代学习观测器的姿态容错控制方法。该方法通过设计迭代学习观测器,以较小的计算量实现对执行机构发生的故障以及安装偏差进行精确的估计。并利用观测器的观测结果设计滑模控制器,使处于外部干扰条件下的卫星系统在执行机构发生故障的情况下可以快速稳定地完成姿态机动任务。进一步基于Lyapunov稳定性定理证明了迭代学习观测器及控制器的全局渐近稳定性。针对反作用飞轮存在不确定性及故障的情况下进行仿真,仿真结果表明,与同类容错控制方法相比,所提方法可以更加快速精确地对故障进行估计并完成姿态控制。     

Comparative evaluation of prior versus progressive articulation of preference in bicriterion optimization

Procedures for solving multiple criteria problems are receiving increasing attention. Two major solution approaches are those involving prior articulation and progressive articulation of preference information. A progressive articulation (interactive) optimization approach, called the Paired Comparison Method (PCM) is compared to the prior articulation approach of a priori utility function measurement in a quality control decision environment from the perspective of the decision maker. The three major issues investigated included: (1) the ease of use of each method, (2) the preferences of solutions obtained, and (3) the insight provided by the methodology into the nature and structure of the problem. The problem setting involved management students who were rquired to determine an acceptance sampling plan using both methods. The PCM provided the most preferred solutions and was considered easier to use and understand. The prior articulation of preference method was found to give more insight into the problem structure. The results suggest that a hybrid approach, combining both prior preference assessment and an interactive procedure exploiting the advantages of each, should be employed to solve multiple criteria problems.     

一种具有遗忘特性的在线学习算法框架

基于凸优化中的对偶理论,提出了一种具有遗忘特性的在线学习算法框架。其中,Hinge函数的Fenchel对偶变换是将基本学习问题由批量学习转化为在线学习的关键。新的算法过程是通过以不同方式提升含有约束变量的对偶问题实现的:(1)梯度提升;(2)贪婪提升。回顾了以往的相关研究工作,并指出了与之的区别与联系。人造数据集和真实数据集上的实验结果证实了算法框架的有效性。算法可以很好地处理数据流中的分类面漂移问题,为设计和分析新的在线学习算法提供了一个新的思路。