Font Size: a A A

Study On Machine Learning And Its Applications In Multi-Agent Game Learning

Posted on:2006-07-09Degree:DoctorType:Dissertation
Country:ChinaCandidate:Q D WangFull Text:PDF
GTID:1118360182465736Subject:Systems Engineering
Abstract/Summary:PDF Full Text Request
The theory and techniques in multi-agent system can be used as a novel method to analyze, design and implementation distributed open system. However, with the rapidly development of relational fields, the environment where multi-agent systems operates in becomes more and more large, open, dynamic and uncertain. In order to adapt to complicated environment, it's urgent to introduce learning mechanism into multi-agent system, and build intelligent agent with self-learning ability by techniques from artificial intelligence. Learning task of multi-agent system includes learning knowledge for decision support from data and information accumulated by machine learning methods, and game learning for establishing rules of multi-agent collaboration, correspondence and competition. Therefore, researching machine learning and game learning methods is very important to development of multi-agent system learning.In machine learning field, mass data, complicated data, data with noise and missing value, data scarcity, all bring great challenge to research and design of machine learning methods. How to improve performance, efficiency, convergent speed of machine learning algorithms that deal with mass data, how to generalize existing machine learning methods in order to analyse data with noise and missing value, and study machine learning methods that can do well with small size samples set, have become the problems that needed to be solved urgently. Then, ensemble learning, reinforcement learning, statistic machine learning, and machine learning methods for mass data, have become the researching focus of machine learning field. In game learning field, it's the main researching direction that players learn to play optimal strategy by reinforcement learning and other machine learning methods. The key to study and design game learning methods is how to ensure its rationality and convergence to optimal strategy. Therefore, This dissertation studies main machine learning methods. Several generalized machine learning methods are presented, and their performance are analyzed theoretically and experimentally. At the same time, machine learning methods are applied in game learning. Stochastic game and differential game learning methods based on ensemble learning, reinforcement learning, and other machine learning methods are studied. Game learning algorithms are designed accordingly, and their rationality and convergence are analyzedexperimentally. Main topics in this dissertation include(1) Hierarchical rough set model, conception hierarchy based dataset partition method and association analysis algorithms are presented by introducing conception hierarchy into rough set theory and association analysis. Hierarchical rough set model is a generalization of classical rough set model by introducing measures for depth and extent of knowledge. Hierarchical rough set model's application in accuracy assessment of aerocraft fall points implies that it has more powerful data analysis ability and wider application fields than classical rough set model. Conception hierarchy based dataset partition method cuts the dataset into independent sub-datasets that can be analysed to get as same result as association analysis in whole dataset. It can improve efficiency of association analysis greatly, and association analysis algorithm based on it can be adapted to parallel algorithm easely. Experiments show the efficiency of conception hierarchy based dataset partition method and association analysis algorithms.(2) A new ensemble learning method, attributes reduct based ensemble learning, is presented by introducing rough set theory into ensemble learning. In attributes reducts based ensemble learning, attributes reducts are computed in training dataset with weight distribution on every examples, and weak classifiers are built on attributes reducts. Because weak classifiers built by attributes reducts based ensemble learning have stronger independence on classification error than other ensemble learning methods, the performance of ensemble classifier can improved greatly. Experimental results proved the effectiveness of attributes reduct based ensemble learning methods.(3) A multi-label classifier learning method, support vector machine decision tree, is presented. Support vector machine decision tree is a decision tree that each inner node is a support vector machine and leaf node is classification label. Support vector machine decision tree learning inherits the excellent feature of support vector machine that can get less generalization error when number of training examples is limited, and decision tree learning that can get high classification accuracy with tree classifier that can be understood by human easily. The problem of measuring hydrophobic grade of insulated material is solved by support vector machine decision tree successfully.(4) Association analysis based hierarchical reinforcement learning is presented. In this hierarchical reinforcement learning, sub-goals are discovered by associationanalysis. In order to improve efficiency and convergence rate of learning, hierarchical Options are constructed from sub-goal discovered which could be used to omit detail of reinforcement learning task. Experimental result shows that the sub-goals discovered, hierarchical Options constructed from sub-goals, and strategy learned among Options, can also be used to accelerate learning in correlative reinforcement learning task as background knowledge. Experimental results proved the effectiveness of association analysis based hierarchical reinforcement learning.(5) Matrix game and stochastic game which can be used to model multi-agent game are introduced. The rational and convergent rules to measure performance of game learning algorithms are constructed. Ensemble learning based stochastic game learning method is presented, and it is proved theoretically and experimentally that this game learning method satisfies the rational and convergent rules.(6) Moreover, application of machine learning methods in game learning is study. Three machine learning based differential game learning methods, variable learning k nearest neighbor learning based differential game learning methods, sub-goal discovering Q reinforcement learning based differential game learning methods, ensemble learning based differential game learning methods, are presented. Testing of three differential game learning methods on pursuit differential game shows that ensemble learning based differential game learning methods is the most robust differential game learning method that satisfies the rational and convergent rules of game learning algorithms.
Keywords/Search Tags:Multi-Agent System, Machine Learning, Game Learning, Rough Set, Association Analysis, Ensemble Learning, Support Vector Machine, Decision Tree, Reinforcement Learning
PDF Full Text Request
Related items