Research On Ensemble Learning Methods Based On Evidential Reasoning Rule

As an important research direction in the field of machine learning,ensemble learning can obtain stronger capabilities than a single learner by integrating multiple base learners.Because of its excellent performance,ensemble learning has not only made remarkable achievements in regression,clustering,classification and other tasks,but also has been widely used in target recognition,target tracking,intrusion detection,auxiliary diagnosis and other practical fields.In ensemble learning,the selection of appropriate base learners and the use of effective combination strategies have an important impact on the results of the final ensemble learning model.At present,the combination strategy in ensemble learning is single,it is difficult to effectively mine the information between basic learners,and there is a lack of effective measurement to evaluate the diversity of basic learners.In the combination of basic learners,all basic learners are often integrated,ignoring the diversity between basic learners,which not only misses the optimal performance that can be obtained by the integration of these basic learners,but also increases the cost of computing operation.In order to solve the above problems and further improve the performance of integrated learning,this paper makes an in-depth study on the integrated learning method,which is mainly divided into the following three aspects:(1)Combination strategy.Aiming at the problem that the voting fusion methods such as average method,which are commonly used in the integration of basic learners in ensemble learning,are difficult to mine the information between basic learners effectively,and the learning methods such as stacking need to spend high time and cost.In this paper,an evidential reasoning(ER)rule based on combined weighting method is proposed as the combination strategy in integrated learning.Aiming at the problem that the traditional evidential reasoning rules cannot set reasonable weight for the base learner,the combined weighting method is proposed to design the evidence weight index considering subjective and objective information for the evidence reasoning rules.In addition,the scope of evidence weight is also discussed in the process of combination.(2)Basic learner selection.Aiming at the existing problems of diversity measurement and evaluation of base learner diversity,which are single,inaccurate and unable to effectively guide the integration of base learner.This paper proposes an updatable fusion(UF)measure to select base learners.This method uses evidential reasoning rules to fuse the diversity measures from different fields,and adds the process of significance test before fusion to test the correlation between the diversity measures used for fusion and avoid the error caused by fusion similarity measures.(3)Prototype implementation.This paper combines the research on the combination strategy and the selection of basic learners in integrated learning and applies it to the field of meteorological recognition.The prototype and several modules of meteorological recognition system based on integrated learning are analyzed and designed.At the same time,it is developed and implemented.The test verifies the feasibility of this research.In this paper,the research on combination strategy and base learner selection can effectively improve the performance of integrated learning model.In the aspect of combination strategy,the combination strategy of evidence reasoning rules based on combination weighting method can provide more reasonable and general weight for the base learner of evidence reasoning rules,and improve the ability of evidence reasoning rules to mine the information of the base learner.In the aspect of base learner selection,the proposed base learner selection method based on updatable fusion metric can more accurately evaluate the diversity of base learners and obtain a better combination of base learners.Finally,the proposed combination strategy is combined with the base learner selection method,and the prototype of meteorological recognition system based on integrated learning method is constructed,which further verifies the effectiveness of the method proposed in this paper.