Research On Data-Driven Method Of High Dimensional Black Box Optimization

The optimization problem without explicit mathematical expression of objective function or constraint function is called black box optimization problem.If there are many decision variables involved,it is further classified into the category of high dimensional black box optimization.Since there is no explicit algebraic expression to calculate,the traditional gradient-based optimization algorithm is not suitable for solving this kind of problems.because it only relies on the meta-heuristic optimization mechanism of iteratively executing "generation+test" that is a kind of flexible and open computing framework,intelligent optimization algorithm has become the algorithm tool of choice for solving high-dimensional black box optimization problems.Without loss of generality,"generation" refers to the intelligent optimization algorithm constantly tries trial solutions in the solution space in the way of population interaction,while "test"generally evaluates the quality of "generation" new solutions through calculation,simulation,measurement,experiment and other ways,so as to provide a basis for generating superior solutions and eliminating inferior solutions.As a necessary algorithm tool to solve highdimensional black box optimization problems,intelligent optimization algorithms still face challenges.In the optimization iteration process,a large number of new solutions generated by frequent population interaction of intelligent optimization algorithms must pass "test" to obtain accurate solution quality evaluation.For high-dimensional black-box optimization problems,the"test" process of new solutions usually has a high computational cost,so intelligent optimization algorithms always face extremely high "test" pressure when solving high-dimensional black-box optimization problems.In order to effectively solve this problem,this topic takes the data-driven optimization as a means to treat the real "test" solution at high computational cost in the optimization process as a training sample to train the lightweight surrogate model,and then partially replace the subsequent"test" process with the "prediction" of the agent model’s solution quality.To achieve the purpose of relieving algorithm "test" pressure.In high-dimensional black box problems,the increase of problem dimensions leads to "dimension disaster",which is manifested in that the exponential expansion of solution space makes it difficult to directly construct the agent model,and the difficulty of hitting the optimal solution increases,which puts forward higher requirements for the optimization ability of the algorithm.In this paper,a particle swarm optimization algorithm based on a two-granularity proxy model is proposed.The coarse-grained model is used to describe the low frequency information of the solution space,and the fine-grained model is used to record the high frequency information of the promising solution region.The two cooperate to give a reasonable fitness prediction value.In order to improve the optimization ability of the optimization algorithm,the diversity feedback mechanism is introduced into the particle swarm optimization algorithm to evaluate the population diversity by population entropy,and the inertia weight and learning factor of the algorithm are adjusted by feedback to balance the "exploration"and "exploitation" ability of the algorithm,so as to avoid the algorithm falling into premature convergence.Experiments were carried out on coverage enhancement problems of directed sensor networks and tumor target planning problems,and the experimental results proved that the two-granularity model can guide the algorithm optimization well,and the diversity feedback mechanism is effective for improving the algorithm optimization ability.