Reinforcement learning for DEC-MDPs with changing action sets and partially ordered dependencies