Domain-driven Metrics for Reinforcement Learning: A Case Study on Epidemic Control using Agent-based Simulation