Seems you could equivalently treat rule breaking as a loss, and any algorithm sophisticated enough to learn how to win will also learn to avoid breaking the rules.