Modelling the Dynamic Joint Policy of Teammates with Attention Multi-agent DDPG