A batch of concurrent evaluations are queried for completions and a partial set is returned to the algorithm


  • Alias: None

  • Arguments: None


In the nonblocking case, the set of running evaluations is queried for completion and a partial set of results is returned to an asynchronous algorithm, which can then generate new evaluations based on these partial results.

In all cases, the algorithm steps will be subject to parallel timing variabilities / race conditions such that these runs will not generally be repeatable regardless of seed specification.


When specifying nonblocking:

  • In the context of parallel pattern search, all points in the pattern may not be evaluated. The first improving point found becomes the next iterate.

  • In the context of adaptive refinement of Gaussian process (GP) models, look-aheads are back-filled with new refinement candidates based on the latest updated state of the GP.