Async tell and eval

[louiskirsch]

Tell has high CPU usage but no GPU usage, if my eval is GPU bound I have my GPUs idle during tell.
Is there a possibility to run another ask and eval during the current tell?

[nikohansen]

There is a related issue: CMA-ES/libcmaes#204

Running ask and eval in itself is a matter of implementation (though ask uses the state of the object that tell changes, so I guess you want to call ask before and not during the change). The question then is whether the result of eval is still useful after tell has been executed. There are quite a few caveats one would need to figure out. Using a significant number of samples in tell from a different (not updated) distribution is in particular problematic for step-size control and for controlling unbiasedness in the update. There are implementation details that could be crucial and it is not at all clear to me under which circumstances this would work or fail.