Offline-Online Reinforcement Learning for Synthon Completion in Retrosynthesis

RLSynC, a multi-agent reinforcement learning method, can outperform state-of-the-art synthon completion methods by up to 14.9% through offline training and online data augmentation.