The problem I’m going to assume applies here is the same as why we can’t train self-driving cars with a simulation. It’s a paradox; if the simulation was robust enough to train cars for the real world, then creation of the simulation itself WOULD be the solution since it already knows all the rules and correct reactions.
In other words, I’m sure they got their simulation to work, within the confines of a limited, non-paradoxical, error-prone programming.