Learning a relational policy to solve planning problems by ranking actions using a new graph representation and decoder architecture