A Multi-Agent Policy-Gradient Approach to Network Routing