I imagine you'd at least have to inject domain knowledge into the model, the ruleset isn't as simple as chess or go, and all the places where a bunch of actions

I imagine you'd at least have to inject domain knowledge into the model, the ruleset isn't as simple as chess or go, and all the places where a bunch of actions have to be strung together just so will lead to a very spiky parameter space where gradient descent struggles.