New training approach helps AI agents perform better in uncertain conditions