Study On Intelligent Autonomous Optimization Mechanisms Based-on Dynamic Bayesian Networks

In the dissertation, autonomous optimization based-on Dynamic Bayesian Networks(DBN)is studied. The research include many works. The first, DBN inference algorithm is discussed. Discrete hide variable DBN inference algorithm and continuous hide variable DBN inference algorithm have been studied, algorithm complex and useable range are analyzed. Secondly, constant DBN structure learning model in smooth random system and changing DBN structure learning model in unsmooth random system are investigated. Thirdly, DBN structure optimization algorithm is researched. DBN structure measurement is studied and structure measurement can be decomposed is put forward. Lastly, a new optimization algorithm based on evolutionary algorithm(EA) and DBN is discussed. DBN inference and structure learning with BN decision-making diagram have been combined to control agent. Simulation results are provided to demonstrate the effectiveness of the proposed algorithm.The main contributions of the dissertation are as follows.(1) The dissertation proposes some reasonable modes to choose DBN inference algorithm in various environment and we do some experiments to testify. Three DBN inference algorithms are contrasted and some conclusions are gained. The three algorithms include using BN inference algorithms to deal with DBN inference without time factor, translating DBN into HMM and using Viterbi decode algorithm to infer, using ready DBN inference algorithm to infer. Through experiments, three principles about DBN inference are gained. The firstly, when there are few networks variables or DBN's coupling is strong, we should choose HMM inference machine to infer. Secondly, when there are more networks variables or DBN's coupling is feebleness, we should use ready DBN inference machine to infer. Thirdly, we should use LDS inference algorithm to infer KFM and its derive model. Those conclusions are basis of using DBN inference to apperceive environment.(2) On the basis of DBN measurement decompose, the dissertation proposes measurement decompose can reduce DBN structure optimization time and replant BN structure learning algorithm to DBN, some experiments are done to testify. The firstly, extend BD and BIC measurement to DBN and we gain a integrated decompose formula. Secondly, many simulations are done to test DBN measurement decompose structure learning and validate ideas are right. The conclusions are basis of DBN structure learning model.(3) Learning DBN structure based on BOA is proposed to optimize DBN structure optimization. The algorithm include four steps, 1. Establishing a fine solution group S(t): using all kinds of choice mechanisms to select S(t) from current population. 2. Finding a DBN to matching S(t): using BD metric to find the better network structure according as S(t). 3. Producing a new choice solution: using joint probability distributing of network diagram to produce a new individual group. 4. Creating next population: using new individual group to replace elderly group some chromosome and update population as newer population. The four steps are carried out repetitively to meet the termination rule. The second step is important, we present GA based on greed algorithm(GA-GS) to deal with that. The DBN optimization is key sector to apperceive environment.(4) Constant DBN structure learning model in smooth random system and changing DBN structure learning model in unsmooth random system are proposed. The Fuzzy self-adapt measure algorithm is presented to dynamic decompose unsmooth random process. The algorithm is key technique to dynamic apperceive environment.(5) A new autonomous optimization mix-optimization model based on DBN and EA is proposed. The idea use Dynamic Bayesian Network as a transfer networks of evolutionary algorithm from t to t+1 generation. Through Dynamic Bayesian Network, we can advance a static state optimization way based on Bayesian optimization algorithm and genetic algorithms about probability model to dynamic optimization. Dynamic Bayesian Networks can change optimization basic conditions and establish optimization direction through perception of situation change. It can guidance agent achieves a series of complex tasks without supportment of person.