Decentralized Multi-Agent Reinforcement Learning with Global State Prediction