Research On Heuristic Optimization Algorithms For Large Scale Global Optimization Problems

The heuristic optimization algorithm is a kind of optimization algorithm which simulates and utilizes laws of nature or biological habits.Compared with traditional algorithms,heuristic optimization algorithms have the advantages of stronger robustness,versatility and ease of use when solving optimization problems.In recent years,the research of heuristic algorithms is widely concerned.However,when the heuristic optimization algorithms tackle with the large scale global optimization problems(LSGO),the performance deteriorates sharply,plaguing its further application.The main reasons for this problem are that the search space of the optimization problem is growing exponentially,and the redundant dimensional information seriously interferes the optimization algorithms’ performance with the number of decision variables increasing.Therefore,aiming at the problem,this thesis studies to improve the performance of heuristic optimization algorithms in LSGO problems.The main research contents and contributions of this thesis are as follows:Firstly,an optimization framework that combines manifold learning and heuristic optimization algorithm is proposed,mining the low-dimensional intrinsic information embedded in high-dimensional optimization particles by manifold learning.In this framework,we propose a new and portable mechanism,which approximately achieves the feature mapping of low-dimensional manifold space to original high-dimensional space,so as to extract the effective dimensions information affecting the current movement of the particles in the original highdimensional space.Then,the method of dividing dimensions into subgroups and dimensional influence coefficient are designed.So,by increasing the effect of effective dimensions and reducing the interference of redundant dimensions on the particle motion,the optimized particles could be guided to move in a more suitable direction to achieve better optimization performance.Secondly,a Manifold-Guided Gravitational Search Algorithm(MGGSA)is proposed,combining the Gravitational Search Algorithm(GSA)and manifold learning according to the optimization framework proposed in this thesis.In MGGSA,the method of dividing dimensions into subgroups and dimensional influence coefficient are designed in detail.According to the characteristics of the iterative update mechanism in GSA,the dimensional influence coefficient is applied to gravitational force of particles to change the magnitude of the gravitational force and the direction of the resultant force,so as to guide the particles to move in a more suitable direction.By introducing manifold learning,the performance of GSA to solve LSGO problems is effectively improved.Finally,a Manifold-Guided Particle Swarm Optimization(MGPSO)is proposed,combining the Particle Swarm Optimization(GSA)and manifold learning according to the optimization framework proposed in this thesis.In MGPSO,according to the characteristics of the iterative update mechanism in PSO,the dimensional influence coefficient is applied to the velocities update of the particles.So,the particles cloud be guided to move in a more suitable direction by changing the magnitude of the particle velocities and the direction of the resultant velocities.By introducing manifold learning,the performance of PSO to solve LSGO problems is effectively improved.