Does anyone know of a variant of the CMA algorithm that is suitable for an asynchronous parallel implementation?
The conventional version of the algorithm allows one to evaluate the fitness function at different points in parallel, but one has to wait until all computations are finished before starting the next iteration. If the computational cost of the fitness function is heterogeneous, the parallel efficiency quickly degrades. Ideally an asynchronous version should be such that parameters are not disadvantaged when the fitness function takes longer to compute.