On discovery and learning of models with predictive representations of state for agents with continuous actions and observations