Does It Pay to Extend the Perimeter of a World Model?

Will the cost for observing additional real-world phenomena in a world model be recovered by the resulting increase in the quality of the implementations based on the model? We address the quest for optimal models in light of industrial practices in systems engineering, where the development of control strategies is based on combined models of a system and its environment. We introduce the notion of remorsefree dominance between strategies, where one strategy is preferred over another if it outperforms the other strategy in comparable situations, even if neither strategy is guaranteed to achieve all objectives. We call a world model optimal if it is sufficiently precise to allow for a remorsefree dominating strategy that is guaranteed to remain dominant even if the world model is refined. We present algorithms for the automatic verification and synthesis of dominant strategies, based on tree automata constructions from reactive synthesis.