Learning Generalized Policy Automata for Relational Stochastic Shortest Path Problems