Evolutionary Many-Task Optimization Framework Based on Machine Learning
Evolutionary multitasking optimization is one of the research hotspots in the field of computational intelligence in recent years.Its principle is to enhance the efficiency of evolutionary algorithms to solve multiple tasks simultaneously through knowledge transfer between tasks.Since inter-task similarity plays an important role in promoting the positive knowledge transfer between tasks,how to measure the similarity between tasks has become one of the key research directions.At present,the existing evolutionary multitasking algorithms can be divided into evolutionary multi-task algorithms and evolutionary many-task algorithms according to the number of tasks processed.However,these algorithms are not efficient enough in handling both multi task and multi task problems simultaneously.For example,when evolutionary multitasking optimization tackles two tasks,the selection of auxiliary task is limited to one of them,and can lead to the lack of flexibility of knowledge transfer between tasks in the aspect of dealing with many-task.In addition,the dynamic nature of task knowledge transfer should occur across tasks,rather than limited to a certain task or several tasks,which requires the algorithm to have the ability to adaptively find other individuals similar to the target task individuals in iteration.Based on machine learning,this paper proposes a framework for solving evolutionary multitasking optimization,named as MaTML,which combines all task-associated subpopulations to form a unified initial populations.MaTML employs the skill factors of the target task and its corresponding population individuals to construct labels and training sets respectively,utilizes the 10-fold cross-validation to fit the model,and then applies the model to predict that population individuals similar to the target task compose auxiliary populations,so as to achieve more positive knowledge transfer in evolutionary optimization.Specifically,in each iteration,the skill factor of the target task is taken as the training label,and all the individuals corresponding to the label,that is,the target task individuals,are combined to form the training sample.Then the model is trained based on the machine learning algorithm,and the model is applied to predict the individuals in the population except the target task individuals.Since the characteristics of the individuals correctly predicted by the algorithm are essentially similar to those of the target task individuals,these individuals can be labeled as auxiliary individuals of the target task.The proposed algorithm can find the auxiliary populations of the target task in the dynamic population individuals,so it can not only flexibly select similar auxiliary tasks for three or more tasks,but also solve the problem of effectively selecting auxiliary tasks at that time the number of tasks is two.What's more,the selection of population individuals for auxiliary tasks of the target task is variable and dynamically adaptive in MaTML,and there is no need to associate a subpopulation with the target task.Compared with state-of-the-art multi-task algorithms and many-task algorithms in CEC2017 test suite and WCCI2020SO test suite respectively,the experimental results show that MaTML has superior or competitive performance in optimizing multi-task or many-task problems.We also conducted a detailed study on MaTML's computing resources,model performance,model stability,and related components.Finally,an optimization problem of real-world was served as test to further verify the effectiveness of MaTML.
NETL
NSTL
万方数据
