Quantifying Agent Interaction in Multi-agent Reinforcement Learning for Cost-efficient Generalization