Multi-agent Actor-Critic Reinforcement Learning Based In-network Load Balance