Learning to Coordinate with Humans using Action Features