Improving Generalization of Reinforcement Learning for Multi-agent Combating Games