If you run REPR on YACHT for 6 iterations, you randomly get somewhat different answers from ERMA on the last iteration

[jeckstei]

First, there should only be one optimal answer.

Second, I think RMA should be deterministic when run in serial (even if these means fixing some random number seed)