IJCAI (2025)
Deunsol Yoon, Sunghoon Hong, Whiyoung Jung, Kanghoon Lee, Woohyung Lim
Abstract
Naphtha Cracking Center scheduling aims to develop optimal multi-week plans under operational constraints and fluctuating demand. Our prior work (Hong et al., 2024b) introduced a multi-agent reinforcement learning (RL) system that is currently deployed in a petrochemical plant. However, standalone RL agents face several limitations: the environment is sensitive—one suboptimal action can invalidate the entire plan—and reward functions are often difficult to specify. We propose Population-Based Multi-Scenario Planning (PBMSP), a novel planning algorithm designed to complement RL agents. PBMSP maintains a diverse set of candidate schedules optimized for distinct objectives and constraints, and extends RL-based scheduling by enhancing adaptability, stability, and operational profitability.