Scale high-performing ad sets without restarting the learning phase? This could be extremely helpful.