Alongside large language models, autonomous driving and humanoid robots, AI-driven optimization of industrial manufacturing has emerged as a major application of AI. In 2024, Kim Jeong-hye, a PhD student of Professor Youngchul Sung, interned on LG AI Research’s reinforcement-learning team, where she tackled a range of process-optimization challenges across LG Group’s production facilities.
That team applied optimization majorly based on reinforcement learning to LG Chem’s Daesan plant’s naphtha-cracking facility (NCC), improving production efficiency by 3%, far beyond of 0.1% of initial expectation, and yielding an extra KRW 10 billion in annual profit for that plant alone. Because training reinforcement learning agents via on-line interaction in a production environment is impractical, such optimization typically relies on offline reinforcement learning, which optimizes policies with pre-collected data.
Jeonghye contributed to the development of PARS, a novel offline reinforcement learning algorithm that significantly outperforms existing methods. By enhancing the neural network’s feature resolution with reward scaling with layer normalization, this new approach better differentiates between in-sample and out-of-distribution data, eliminating Q-value divergence, the core issue of off-line reinforcement learning. This advancement promises to accelerate future production-process optimizations as well as many RL applications with difficulty in on-line environment interaction.
This research result will be presented as a Spotlight paper at the International Conference on Machine Learning (ICML) 2025.
Related Yonhap News article: https://www.yna.co.kr/view/AKR20250613153400003
