Parallel optimizer variants are implemented somewhat differently from their sequential versions. The typical way of parallelizing a multi-agent optimizer is to maintain and update the population of agents on one execution thread and then distribute only the computation of the fitness to multiple execution threads. This makes it easier to synchronize access to the data. However, this also means the entire population must be processed before improvements can become effective and be used in the computation of new candidate solutions. This changes the dynamic behavior of the optimizer and means it requires different behavioral parameters to work well, which may not necessarily work as well as the optimizer's sequential version.