يادگيري مشاركتي بر مبناي خبرگي چند معياره در سيستم هاي چند عامله

چكيده انگليسي :

84 Multi Criteria Expertness based Cooperative Learning in Multi Agent Systems Esmat Pakizeh Hajyyar e pakizehhajyyar@ec iut ac ir Date of Submission 2011 04 13 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Maziyar Palhang palhang@cc iut ac irAdvisor Mir Mohsen Pedram pedram@tmu ac irAbstractSince cooperation is the key to success in most biological and artificial communities the capability ofcooperation in multi agent systems is critical in achieving better solutions Multi agent cooperative learningresults in higher efficiency and faster learning compared to individual learning due to more resources ofknowledge and information Better cooperative strategies may speed up and improve learning Cooperative learning is a group learning activity organized so that learning is dependent on the sociallystructured exchange of information between learners in groups while each learner is held accountable forhis her own learning The aim of this thesis is to contribute to an answer to the question How can agentsenjoy of exchanging information during the cooperative learning process in order to achieve betterindividual and overall system performances Researches in cooperative learning showed that the question isnot only what type of information to exchange but also how to use shared information Nowadays the majority of researches in multi agent cooperative learning field focus on ReinforcementLearning RL as their basic learning method RL is one of the most prominent machine learning methodsdue to its unsupervised learning structure and continuous learning ability even in a dynamic operatingenvironment Applying this learning to cooperative multi agent systems not only allows each individualagent to learn from its own experience but also offers the opportunity for the individual agents to learn fromother agents in the system so that the speed of learning can be accelerated During the life cycle human learns through different experiences over different time periods of his life Sometimes the experience is quite successful and sometimes it completely fails Individual s character isformed based on all of the gained experiences whether they are good or bad Everyone will make hisdecisions based on his formed character This is what we have attempted to translate into the realm of multi agent systems learning For this purpose in our study a novel concept named Multi Criteria Expertness isintroduced that takes advantages of worthy information about different experiences of agents in a cooperativemulti agent system In addition in this thesis a new cooperative learning algorithm is proposed which enjoys of multi criteriaexpertness concept and attempts to cooperate more efficiently The proposed method has high ability to usemore knowledge and information compared to existing methods which leads to high performance Most ofrelated researches in the cooperative learning field have concentrated on improving the group learning bychanging the agents Q tables through knowledge transferring while in the proposed algorithm transferredknowledge is used as a clue in action selection of all the agents In other words each agent keeps its partialknowledge and is guided through collective knowledge that appears in the cooperative Q table Thisinnovation made agent s conductivity more stable Consequently a suitable cooperative Q table speeds up thelearning process and if the cooperative Q table is not suitable the agent s Q table is capable of compensatingits incompetence Experimental results performed on a sample maze world and a hunter prey environment show the highpotential of proposed algorithm in producing better cooperative strategy Keywords Multi Agent Systems Cooperative Learning Multi Criteria Expertness ReinforcementLearning Knowledge Transfer

پاكيزه حاجي يار، عصمت

يادگيري مشاركتي بر مبناي خبرگي چند معياره در سيستم هاي چند عامله

يادگيري تقويتي , انتقال دانش