Exploration with Unreliable Intrinsic Reward in Multi-Agent Reinforcement Learning