Can the population size be adjusted on the fly?

[Gabriel-p]

A larger population size results in a larger ESS, but also in a larger probability of getting stuck in a local minima. A smaller population size better explores the parameter space and (at least in my case) usually ends up finding better solutions, but at the cost of a smaller final ESS.

My ideal set up would be to begin with a population_size=50 abd increase it by a fixed number, say 10, at each new generation; up to a maximum of say 200.

Can this, or a similar strategy, be implemented in pyABC? The closest I can think of is to perform several small runs, at each run manually increasing the population_size and adjusting the priors' range using the distributions found by the previous run.

[arrjon]

There exists the pyabc.populationstrategy module, where you can define your strategy. Currently, you can either use a fixed list of population sizes or an adaptive strategy. You can give the stratgy to population_size instead of a constant size. Is that what you were looking for?

[Gabriel-p]

Thank you @arrjon, yes that is what I was thinking about. Is there a way to tell ListPopulationSize() to just use the latest value in the list when the time t increases above the length of the list? I don't know how many generations the code will run for (I use max_walltime) and when the strategy goes beyond the end of the list it throws and IndexError.

[arrjon]

Unfortunatley, there is not. You need to provide a list as long as the number of possible populations. However, you can set the max_nr_populations argument in abc.run to be sure not to exceed the list.