Transfer Reinforcement Learning for Differing Action Spaces via Q-Network Representations