Distributed Value Function Approximation for Collaborative Multi-Agent Reinforcement Learning