[精选]适性化多代理人网际网路环境资讯侦搜

《[精选]适性化多代理人网际网路环境资讯侦搜》由会员分享，可在线阅读，更多相关《[精选]适性化多代理人网际网路环境资讯侦搜（21页珍藏版）》请在装配图网上搜索。

1、適性化多代理人網際網路環境資訊偵搜適性化多代理人網際網路環境資訊偵搜 Collaborative Multiagent Adaptation for Business Environmental Scanning through the Internet 劉瑞瓏Rey-Long Liu中華大學資訊管理系中華民國 92 年 11 月 18 日1Outline Introduction Business Environmental Scanning through the Internet(ES)User-Centered,Continuous,and Resource-Bounded ES(UC

2、RES):Goal&challenges Intelligent Multiagent Technology Multiagent Adaptation for UCRES Overview of AESA The Scanning Agents The Controlling Agent Experiment Conclusion2Introduction3 UCRES:User-centered,Continuous,and Resource-bounded ES Goal&clallenges Finding IOI continuously Timeliness(TES):minimi

3、zing the average time delay of finding IOI Completeness(CES):maximizing the percentage of IOI found Controlling the resource consumed Effectiveness(EES):maximizing the possibility of finding IOI in each inquiry of data Considering the preferences of the user(i.e.adding importance weights to TES,CES,

4、and EES)Weighted TES Weighted CES Weighted EES4 Related multiagent technology for UCRES Information gathering agents Aiming to satisfy users one-shot needs,e.g.Intelligently locating the information with the additional consideration of time/cost/quality tradeoffs,and Filter out irrelevant informatio

5、n But when and how frequently to scan for IOI?Information monitoring agents Aiming to monitor a predefined set of targets periodically or adaptively But where to find the IOI to monitor?5 Adaptive agents for decision support Aiming to provide tailored information for supporting decision making Adapt

6、ing to what to monitor by observing the users preference and/or problem solving strategies But when and how frequently to scan for IOI?Multiagent coordination&bidding Aiming to resolve conflicts and build consensus among the agents But how to define a coordination and/or bidding protocol for UCRES?6

7、 Learning for multiagent coordination,bidding,and organization Aiming to learn How the actions affect each other,What information is required for coordination,When to trigger agent coordination,Usage of what the agents are bidding for,Restructuring of the organization for coordination and collaborat

8、ion But how to simultaneously adapt to Users preference IOIs distribution in the Internet Update behavior of the IOI Limited amount of resource7Multiagent Adaptation for UCRES AESA:Adaptive ES Agents A scalable society of autonomous agents A set of scanning agents A controlling agent Basic rules of

9、the agent society Collaborating for satisfying the users information needs Sharing the limited amount of resource Adaptation to Users preference IOIs distribution in the Internet Update behavior of the IOI Limited amount of resource8Overview of AESASystem AdministratorNotificationEvaluateScanning Ag

10、ent3Controlling AgentTerminateAdjustResourceAESASpawnScanning Agent1Scanning Agent2Scanning AgentnInformation PreferencesResource LimitUser1User2UsermMonitoringDiscoveryIntranetThe WWW InformationSpace9Behavior of each Scanning Agent(1)Goal=Set of information preferences to be satisfied by the agent

11、;(2)Target=Target web site from which the agent is delegated to find IOI;(3)ResourceU=Upper bound of resource consumption by the agent;Repeat (4)If the agent is ready to retrieve information(according to ResourceU),(4.1)Retrieve information from Target;(4.2)If new information is found,(4.2.1)Issue a

12、 Type-E request to get(DegreeOfSatisfaction,SatisfiedGoal);(4.2.2)Aliveness=Aliveness*+DegreeOfSatisfaction;(4.2.3)Trigger suitable procedures,including event logging and user notification;(4.2.4)If Aliveness ,for each target d found in the new information,(4.2.4.1)Issue a Type-S request to spawn a

13、new agent (4.3)If the agent succeeds two times,issue a Type-R-More request for more resource;(4.4)If the agent fails two times,issue a Type-R-Less request for releasing resource;(5)If an information preference is no longer valid,remove the need from Goal;Until Aliveness or Goal is empty;(6)Terminate

14、 the agent by issuing a Type-T request.10Behavior of the Controlling Agent(1)=Upper bound of the resource that may be consumed by the system;Repeat (2)If a user enters a new preference e,generate an agent(if necessary and possible);(3)If a user removes an information preference e,inform all agents t

15、o remove e;(4)If there is a Type-R-Less request from agent k,(4.1)Release(from k)1 percents of resource;(5)If there a Type-R-More request from agent k,(5.1)Allocate(to k)2 percents of additional resource(by agent aliveness);(6)If there is a Type-E request from agent k,(6.1)Reply k with DegreeOfSatis

16、faction=(DOSe*Importancee),for each goal e;(7)If there is a Type-S request to spawn a new agent,(7.1)Reply the requesting agent with (if possible);(8)If there is a Type-T request from agent k,terminate k and release all resource of k;(9)If there is no response from k for a long time,terminate k and

17、release all resource of k;Until the system is terminated.11ExperimentThree dimensions of the update behavior of a web site w whose central category is c:(1)When to update:following exponential updateThe probability of updating w in time interval x follows the probability density function f(x)=e-x,wh

18、ere is the average update frequency of w.(2)Which part to update:(2a)Main content is updated with a probability of 0.5.(2b)Embedded hyperlinks are updated with a probability of 0.5.(3)Way to update:preserving information relatedness(3a)New main content is still of category c.(3b)Let s be the semanti

19、c difference between a new hyperlink and c:(3b1)60%new hyperlinks link to those having s 1.(3b2)30%new hyperlinks link to those having 2 s 3.(3b3)10%new hyperlinks link to those having s 4.Simulating the dynamic information space12Simulating users dynamic preferencesThe importance of an information

20、preference may be dynamic due to various factors in business administration(e.g.evolutions of internal strategic goals,sales seasons)and external environmental eventsAmong the 100 possible categories of interest,10 categories were randomly selected as users information preferences to be satisfied.Ea

21、ch preference k was initially associated with a random importance level Ik(0 Ik 1),which could be dynamically changed.The timing of changing Ik followed the exponential probability density function f(x)=e-x as well,where was the average frequency of changing Ik.For each preference k,randomly ranged

22、from 1/3000 to 1/2000(time/seconds).13 The systems evaluated AESA Basic setting For the controlling agent:1=2=20,=1 inquiry per second,=1 inquiry per 200 seconds Variants AESA-1 For the scanning agents,=0.5,=0.5,and=0.3 AESA-2 For the scanning agents,=0.6,=0.6,and=0.3 Recall that,is related to agent

23、 cloning,is related to agent aliveness evaluation,and is related to agent termination.AESA-1 has a more dynamic agent society than AESA-2(the agents are easier to be generated and terminated)14 BestFirstD Aiming to represent the way of employing best-first discovery to UCRES It was allowed to contin

24、uously traverse through the information space(rather than performing one-shot traversal)It preferred the most promising hyperlinks when traversing the information space To avoid duplicated traversals,if there were multiple attempts trying to traverse to a web site w for a user interest,only one trav

25、ersal was conducted Variants BestFirstD-1 and BestFirstD-2,which selected those hyperlinks whose semantic differences with the user interest were less than or equal to 1 and 2,respectively.Obviously BestFirstD-2 could traverse a larger space than BestFirstD-1.15 BestDPeriodicalM Aiming to represent

26、the way of integrating information discovery with information monitoring It integrated best-first discovery(i.e.BestFirstD)with periodical monitoring,which was the most popular monitoring technique in environmental scanning Once a web site is discovered,BestDPeriodicalM determined a fixed monitoring

27、 frequency randomly ranging from 1/10 to 1/1000 query/seconds(this was a“mercy”to BestDPeriodicalM)Variants BestDPeriodicalM-1 and BestDPeriodicalM-2,which selected those hyperlinks whose semantic differences with the user interest were less than or equal to 1 and 2,respectively Obviously BestDPerio

28、dicalM-2 could traverse a larger space than BestDPeriodicalM-1.16Results(average of 30 runs)00.010.020.030.040.050.060.070.0818054090012601620198023402700306034203780414045004860522055805940630066607020Time StampTimelinessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1BestDPeriodicalM-200.050

29、.10.150.20.250.30.3518054090012601620198023402700306034203780414045004860522055805940630066607020Time StampEffectivenessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1BestDPeriodicalM-200.050.10.150.20.250.30.3518054090012601620198023402700306034203780414045004860522055805940630066607020Time

30、StampCompletenessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1BestDPeriodicalM-217Results(average of 30 runs)00.050.10.150.218054090012601620198023402700306034203780414045004860522055805940630066607020Time StampWeighted effectivenessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1Best

31、DPeriodicalM-200.10.20.30.40.518054090012601620198023402700306034203780414045004860522055805940630066607020Time StampWeighted completenessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1BestDPeriodicalM-200.010.020.030.040.05180540900126016201980234027003060342037804140450048605220558059406300

32、66607020Time StampWeighted timelinessAESA-1AESA-2BestFirstD-1BestFirstD-2BestDPeriodicalM-1BestDPeriodicalM-218Improvements TES:191%WTES:252%EES:1220%WEES:1586%CES:1157%WCES:1413%19Conclusion Finding more IOI in a timelier manner using less resource Collaborative Multiagent Adaptation Multiagent:Scalable ES Collaborative:Sharable resource and findings Adaptation:Adjustable ES for dynamic Information spaces(distribution&update behaviors of IOI),User preferences,and Resource upper bounds20