| Evolutionary algorithm has solved many problems in real life and industry,but it requires a large number of fitness evaluations.In some real optimization problems,evaluating candidate solutions require a lot of time or computing resources,such as physical experiments,a more extreme situation is that no real calculations can be performed on the candidate solutions during the optimization process.In this case,the evolutionary algorithm can only use the existing historical data to guide the optimization,this is called offline data-driven evolutionary optimization.In order to reduce or replace the real fitness evaluation,a common method is to build a surrogate model,and use the model to predict fitness to guide the optimization process.However,there are still some problems that have not been solved.The dimension of the real optimization problem may be very high,and there are many local optimal solutions or constraints,etc.The performance of the algorithm is not satisfactory.In order to solve these problems,the main work of this thesis is as follows:(1)An offline data-driven optimization algorithm based on stochastic ranking and multikernel learning is proposed.The algorithm selects four different radial basis functions as the kernel functions to build radial basis function networks.In order to balance them,the selection operator in the evolutionary algorithm uses a mechanism called stochastic ranking.The algorithm performs better than existing algorithms on high-dimensional problems,and has competitive result even on problems with up to 500 decision variables.The algorithm improves the performance of airfoil aerodynamic design optimization by about 10% compared with the original benchmark.(2)An offline data-driven optimization algorithm based on tri-training in semi-supervised learning is proposed.Existing algorithms are not sufficiently using historical data.The algorithm is difficult to obtain an acceptable solution for high-dimensional problem.To solve this problem,this thesis combines semi-supervised learning with optimization algorithms,and proposes a model management strategy based on tri-training.In each generation,three surrogate models are trained,high-confidence data is selected as pseudo-label data and the three models are updated.Compared with the existing offline data-driven optimization algorithms,this algorithm outperforms other three algorithms on medium-high dimensional optimization problems.(3)The existing algorithms do not consider the offline optimization problem with constraints.If the constraints of the evaluation problem are not expensive,the original algorithm can also solve it.Otherwise,the surrogate model must be used to evaluate constraints.This thesis discusses several commonly used constraint handling techniques in evolutionary algorithms,and designs algorithms combining them with offline data-driven evolutionary optimization.Four test functions with constraints are defined to test different constraint handling methods.The results show that the algorithms based on multi-objective are easier to converge to feasible solutions,while algorithms based on stochastic ranking can get better solutions on the objective function. |
