First of all, I refer to the blog post of https://blog.csdn.net/niuyongjie/article/details/1569671 in this column. I feel that I understand it well, but due to my personal reasons, the relationship of time will be summarized as follows. , if there is a chance to pick up the research in the future, I will update and correct it. If there is any misunderstanding or need to be added, please correct me, thank you.
Finally, what I have to think about is how can I use this algorithm in practice and apply it to specific scenarios?
The function of the optimal solution we need to find is written in the fitness function, such as y=1-cos(3*x)*exp(-x) , he has a solution, so my particle is one-dimensional, we The information that also needs to be provided is the definition domain of this function. Some functions have the most value within the definition domain, so we also need to provide the range of particle motion, so that his random changes and certain rules will not exceed this definition domain, Beyond the domain of his range of value is meaningless. If y=1-cos(3*x1)*exp(-x2) , then he has two solutions of x1 and x2, then the particle at this time is two-dimensional, and the motion range of each dimension must be given out.
In this way, the particle will make adjustments based on its own history and its "circle" or the global particle comparison, and it has weights to adapt to various scenarios.
So in our practice, where does the function come from? And the optimal solution must be the most valuable? I'll save it for later discussion.
After my review, the particle swarm filtering method is used in the detection of moving targets. The basic idea is also similar. He compares the similarity between the histogram of the region of interest ( equivalent to the above function ) and the histogram of the region where the particle is located. The higher the similarity, the greater the weight, and the faster the particle moves to the region of interest. Filter out the ones that are close to 0, add new particles, and the new particles are the copies of the highest similarity, but there will be a problem in the future, that is, the diversity of particles will be reduced, resulting in the final result that may not be the global optimal solution. , but the local optimal solution, that is, the extreme point rather than the maximum point.
Let’s write this for a while, I’m a little sleepy. . . .