1、主体仿真决策框架与应用贺舟 中国科学院大学 经济与管理学院 副教授 博士生导师 www.AgentL|简介|p贺舟l中国科学院大学经济与管理学院，副教授，博士生导师l研究方向：复杂系统建模与仿真、仿真决策l学术兼职：管理科学与工程学会理事、中国运筹学会决策分会理事l社会兼职：工业互联网产业联盟供应链特设组副主席l欢迎报考博士和硕士研究生，也欢迎本科生前来实习l课题组常年招收博士后，若入选中国科学院特别研究助理税前基本年薪不低于40万元，欢迎应聘201主体仿真概述02主体仿真与智能决策03主体仿真决策框架04主体仿真决策应用目录 CONTENT|01主体仿真概述什么是仿真？|p各种定义l用能代表

2、所研究的系统的模型做实验(Korn,1978)l所有支持模型建立与模型分析的活动即为仿真活动(Spriet,1982)l是一种利用模型开展模拟试验研究的科学活动(王维平 等,1997)p共同点l仿真是基于模型的活动5仿真的要素与分类|p三要素l系统、模型、计算机（试验）p分类l离散事件仿真(discrete event simulation,DES)l系统动力学(system dynamics,SD)l主体建模与仿真(agent-based modeling and simulation,ABMS)6仿真的要素与分类|7系统状态可在任意时间点获得系统状态只能在离散时间点获得系统状态随时间连续变

3、化系统状态变量的变化只在特定时刻变化，而在两个特定时刻之间保持不变常微分方程偏微分方程(SD)离散事件系统(DES,ABM)差分方程主体建模与仿真（ABMS）8p主体建模（agent-based modeling，ABM）l模拟主体（agent）的行为和交互，以评估对整个系统的影响p主体仿真（agent-based simulation，ABS）l利用主体模型（agent-based model,ABM）开展模拟试验研究的科学活动p接下来介绍三个较为著名的主体模型|生命游戏9pGame of life（John H.Conway，1970）l模拟生命活动中的生存、灭绝、竞争等复杂现象l主体占据

4、一个方格，有两种状态（生、死）和8个邻居-在当前时刻，如果一个元胞状态为“生”，且八个相邻元胞中有两个或三个的状态为“生”，则在下一时刻该元胞继续保持为“生”，否则“死”去（太拥挤或太孤独）；-在当前时刻，如果一个元胞状态为“死”，且八个相邻元胞中正好有三个为“生”，则该元胞在下一时刻“复活”，否则保持为“死”|生命游戏10滑翔机glider播种机在线网站：https:/bitstorm.org/gameoflife/|生命游戏11大型系统携带内啡肽的蛋白质涌现：系统表现出“总体超过了各部分总和”的特征告知微观规则，难以预见宏观现象；告知宏观现象，难以设计微观规则|种族隔离模型Segregat

5、ion model12当某一居民周围的8个居民中与其相同的比例低于某一阈值，该居民就会“搬家”，随机前往网格中任意空格子Thomas C.Schelling（2005年获得诺贝尔奖）1971年提出：如何解释种族隔离居住的社会现象？|鸟群模型Boids模型13Craig Reynolds（3D电脑动画学者）1987年提出：鸟群行为是如何产生的？电影场景蝙蝠侠归来 Batman Returns(1992)|鸟群模型Boids模型14鸟只有三个规则Alignment：和视野内的鸟群保持整体的飞翔方向Separation：不要和其它鸟靠的太近Cohesion：飞向视野内鸟群的中心|02主体仿真与智能决

6、策智能决策与仿真决策16p我和其它教师正在共同编写决策分析教材l智能决策章节有四章：13.数据驱动的决策；14.算法决策；15.仿真决策；16.决策支持系统p什么是仿真决策？l采用了计算机仿真技术的决策l仿真技术利用了计算机的计算能力，展示和刻画采用不同决策方案后的过程和结果，便于决策者结合其它决策方法制定合理决策|智能决策与仿真决策17p为何采用仿真决策？l狭义的决策就是综合考虑多个方案的结果，从中选择最优或满意解，但从方案到后果，并不是那么明显|买家秀卖家秀主体仿真和决策的结合方式18p1/2主体的行为采用合理的决策方法刻画l1/4反应式行为的建模-代表模型：If-then-else的规则

7、模型、基于概率的规则-适用模型：早期的模型，或者规模庞大的ABM-实际应用：疾病或信息的传播、供应链网络的演化研究Zhao,K.,Zuo,Z.&Blackhurst,J.(2019).Modeling Supply Chain Adaptation for Supply Chain Disruptions:An Empirically Grounded Complex Adaptive Systems Approach.Journal of Operations Management 65(2)190-212.|主体仿真和决策的结合方式19p1/2主体的行为采用合理的决策方法刻画l2/4推理行

8、为的建模-代表模型：BDI模型（Bratman,1987）-适用模型：和人相关的ABM|信念：主体对环境和自身状态的信息的一种评价，可随着时间的推移和环境的变化逐渐积累和更新愿望：主体希望达到的事件状态，可以理解为主体的目标意图：主体决定为达到愿望而实施动作的事件状态，其持续时间小于愿望，也会随着愿望更新而更新主体仿真和决策的结合方式20p1/2主体的行为采用合理的决策方法刻画l2/4推理行为的建模-实际应用：麻将游戏Agent的BDI模型（徐玉朋,2017）-信念：客观资源：已和牌序列、活动牌序列、弃牌序列等【变量】；个性特征：激进度、冷静度、可控度等【随机生成，常数】-愿望：手牌花色小于3

9、种、听普通和牌、听清一色、听对对胡等；以上愿望可以组合，并计算各种目标收益、难度系数等-意图：对当前手牌的价值进行量化，比如有多少个顺子、多少个对子；考虑手牌重新排列组合的可能性，与愿望的目标相对应-动作：出牌、碰牌、明杠、暗杠、和牌等|主体仿真和决策的结合方式21p1/2主体的行为采用合理的决策方法刻画l3/4优化行为的建模-代表模型：NK模型、运筹学模型、启发式算法-适用模型：主体的资源有限从而进行理性决策 资源有限：主体在购买物品时的预算有限，或者决策的选项有限 理性决策：主体的目标明确，即利用有限资源寻求最优的决策|政府企业家庭主体仿真和决策的结合方式22p1/2主体的行为采用合理的决

10、策方法刻画l4/4学习行为的建模-代表模型：神经网络、强化学习-适用模型：高维、高度不确定性|主体仿真和决策的结合方式23p2/2主体仿真的结果作为建模者决策的依据l1/2灵敏度分析，分析单个或多个参数的影响-基于实验设计（experiment design），将一些常数的初始值进行改变，观察ABM仿真结果的变化-可采用方差分析（ANOVA）判断初始值的改变是否对结果产生了显著影响|正交试验方法orthogonal随机拉丁方试验方法random Latin Hypercube最优拉丁方试验方法optimal Latin Hypercube主体仿真和决策的结合方式24p2/2主体仿真的结果作为建

11、模者决策的依据l2/2启发式(多目标)优化-研究者可能需要从理解ABM走向控制ABMl优化ABM的难点-ABM中往往包含大量的随机因素，导致输入的政策变量和输出的政策评估结果之间的映射关系存在随机性-ABM的时间复杂度也随着主体数量的增加而增加，导致ABM的仿真和政策评估更为耗时|主体仿真和决策的结合方式25p2/2主体仿真的结果作为建模者决策的依据l2/2启发式(多目标)优化l主要方法-全局搜索算法-基于仿真的优化算法|03主体仿真决策框架主体仿真决策框架27|宏观决策者宏观决策者微观主体微观主体宏观决策环境微观主体决策新的决策新的决策决策分析决策分析影响多个主体决策行多个主体决策行为的仿真

12、计算为的仿真计算影响影响影响主体对环境的反主体对环境的反馈和决策等行为馈和决策等行为总框架：主体模型+启发式算法多主体模型启发式算法仿真模型输入仿真模型输出适应度(微观反馈)新解(宏观决策)主体决策环境主体感知反馈主体采取决策所有主体决策每个主体收益 参数估计参数估计 运筹优化运筹优化 学习探索学习探索子框架：估计+寻优+学习总框架28|遗传算法主体模型Zhou He,Jie Xiong,Tsan-Sheng Ng,Bo Fan*,and Christine A.Shoemaker(2017).Managing competitive municipal solid waste treatme

13、nt systems:An agent-based approach.European Journal of Operational Research,263(3):1063-1077子框架29|运筹优化（有限资源）参数估计（有限信息）学习探索（有限理性）Zhou He,Chunling Luo,Chin-Hon Tan,Hang Wu,and Bo Fan*(2019).Simulating an agents decision-making process in black-box managerial environment:An estimation-and-optimisation

14、approach.Journal of Simulation,13(2):111-127子框架30|将ESTOPT框架用于仿真主体的定价决策|04主体仿真决策应用Zhou He,Guanghua Han,T.C.E.Cheng,Bo Fan,and Jichang Dong*(2019).Evolutionary location and food quality strategies for restaurants in competitive online food ordering and delivery markets:An agent-based approach.Internat

15、ional Journal of Production Economics,215:61-72外卖市场Motivation32|pOnline food ordering and deliverylOne of the most successful business models in digital economyPartner of this studywaiting time外卖市场Motivation33|pIdentify the key factorslAccording the large-scale surveys conducted in China in 2015 and

16、 2016,diners mainly focused on two factors:Online foodfood quality外卖市场Motivation34|pAgents,interactions and trendswaiting time=food preparing time(food quality)+delivery time(distance from location to diner,delivery policy of the APP platform)1the increasing preferencefor high food qualityshortening

17、 food preparing time23the increasing preferencefor cost savingdiner外卖市场Motivation35|pRQ1:lWhat are the impacts of three possible changes on the food quality and location operations of restaurants,i.e.,-the increasing preference of customers for high food quality,-the shortening food preparing time o

18、f the restaurant,and-the different delivery policies of the online platform?pRQ2:lWhat are the differences between the food quality and location decisions made by the best restaurants and those made by others?外卖市场Assumptions36|pAssumption 1:lDiners and restaurants are represented as discrete points

19、and placed on a two dimensional plane with a polar coordinate system according to their polar coordinates(,)CBD:distance to the CBD center,decision variablerestaurant:randomly generated but fixeddinerCloser to or farther away from the CBD?Diners gather around the CBD(the pole)Randomly generated pola

20、r coordinates 外卖市场Assumptions37|pAssumption 2:lA restaurants food preparation time has a positive and linear relationship with its qualityGood food is worth waiting for.Food quality,decision variable01Food preparation time1 minute20 minutes2外卖市场Assumptions38|pAssumption 3:lWe exclude the other dinin

21、g options like eat-in or order pick-ups at restaurantspAssumption 4:lWe do not consider restaurants that offer delivery servicepAssumption 5:lWhen the diner submits feedback on food quality and waiting time,we assume that submitted food quality always equals the current food quality determined by th

22、e restaurant外卖市场Agents39|pAgent type and model structure外卖市场Agents40|pDiner i at time t:lMore likely to select the restaurant with higher utilityWaiting time from restaurant j to diner iProbability for diner i to choose restaurant jUtility if diner i chooses restaurant jFastestTastiest Logit choice

23、model:better alternatives are chosen more often1Food quality of restaurant j外卖市场Agents41|pPlatform at time t:lDeliver food by solving a complex VRPPDTW-D-Dynamic vehicle routing problem with pick-ups/deliveries and time windows-Orders appear dynamically(no order prediction);pick-up time window start

24、s only when the food is ready;multiple homogeneous riders are traveling with given speed and capacity;-Two objectives:Minimize the maximum waiting time of all dinersMinimize the total travel distance of allthe riders3:weight preference for cost saving :weight preference for user experience外卖市场Agents

25、42|pPlatform at time t:lSuggested by Ele.me,we use the insertion heuristic algorithm to solve the VRPPDTW-D-A rider can picks up multiple takeaway orders at different restaurants-A riders route may change by the algorithm when a new order is received-The real-time delivery scheduling system has to c

26、ontinuously track the location and status of each rider and order外卖市场Agents43|pRestaurant j at time t:lDecide food quality and location to maximize the number of received orders-But the performance is affected by the interweaving decisions of both customers and rivals,as well as the delivery plans g

27、enerated by the online platform-Hence,we incorporate the estimation-and-optimization(ESTOPT)approach proposed by He et al.(2019)to help the restaurant make the joint decisionpRestaurant j at time t:lSo we need to assume a polynomial function,i.e.,-number of received orders is a function of food qual

28、ity and locationlWhy this polynomial form?-Recall that restaurants face a trade-offbetween food quality and waiting time-Partial derivatives are quadratic functions opening downward and the optimal food quality could be within(0,1)Location(r)Food quality(q)number of received orders()外卖市场Agents44|外卖市

29、场Agents45|Table of agent variablesSequential diagram of ABM外卖市场Agents46|pModel validationlBased on real data from ele.me,we simulate a CBD in BeijinglSimulation results are close to real dataModel indicators we are observingA table of validation results can be found in the paper外卖市场Agents47|pThree s

30、cenarios,to answer three research questions1the increasing preferencefor high food quality2shortening food preparing time3the increasing preferencefor cost saving外卖市场Agents48|pThree scenarios,to answer three research questionslScenario A,higher food quality preference-All restaurants increase food q

31、uality;location decisions not affected-Best restaurants make changes more markedly外卖市场Agents49|pThree scenarios,to answer three research questionslScenario B,longer food preparation time-Both decisions are less affected,as the diners are bearing the time cost-Best ones have higher food quality,great

32、er uncertainty in decision-making,and closer to the CBD center外卖市场Agents50|pThree scenarios,to answer three research questionslScenario C,platform tends to save more cost-Location decisions are changed more than food quality decisions-Best ones have much higher food quality and closer to the CBD center非常感谢您的观看|贺舟 中国科学院大学 经济与管理学院 副教授 博士生导师 www.AgentL