|
Hello my dear fellows. I have been thinking on applying Dirichlet processes to Particle Swarm Optimization, haven't found any papers. This means that it either isn't a good idea or no one has thought about that. My idea is to use the DP priors to help maintain the structure of the Overall Particle Swarm, since one of the problems of PSO is that it sometimes looses structure and uses too many particles for the optimization. In basic PSO we usually have a random set of particles that direct themselves to an optimal result based on a velocity and a relative position with the target and the particle closest to the target. Since the particles growing closer to the target are more likely to be eventually the optimal, I was thinking on assigning weights according how many particles they had close to them. (Since every particle would try to grow closer to the optimal one in the group). Thus a particle with a lot of neighbors would be bound to have more neighbors. Thus DP comes into game, if we use a virtual clustering, where the prior is given by something like a Chinese Restaurant Process (An cluster with a lot of elements is bound to attract more elements), we could model this kind of dynamic. My question is: Does this sounds feasible or am I talking bananas here. Thanks |
How would you write down a generative story for this? It seems that the optimization might not necessarily fit in a density. Also, if you don't have conjugacy using the DP gets a lot harder.
Did not think on the conjugacy.
Implement particle assignment as a chinese restaurant process. There is some prior for particle assignments based on whatever the normal process is, then this likelihood is multiplied by the CRP parameters: m_k/(n+alpha-1) if the cluster has members, alpha/(n+alpha-1 otherwise. Note, though, that this tends to result in exponential distribution of cluster sizes, which may not improve your results.
Consider other priors, like a preferential attachment prior (ex. where the "new cluster" term is a fixed fraction) that generates a power-law distribution, or even a distribution that hard-limits particles from joining already-saturated clusters.