Open florianhartig opened 7 years ago
Just some thoughts to the adaptive filtering / likelihood weights.... There are some ways we could do that: a) take the best x% of particles every generation/iteration. We might get into trouble with local optima, since we are completely loosing the exploration part of the filtering step. b) tournament selection c) rank selection. Here we have the problem of choosing probabilities for each rank, without knwoing beforehand how many ranks (particles) there are
Please let me know what you think about that.
At them moment, if iterations = 10, then the likelihood will be divided in 10 equal chunks that are used for the iterative filtering. There was the idea to make the iterations adaptive, such that the interval is always small enough to avoid particle depletion, and large enough so that particles are filtered.