Reinforcement learning (RL) has achieved remarkable success in complex games such as Atari, Go, Pokémon, and Chess, prompting researchers to apply it to real‑world problems like personalized medicine and dynamic treatment regimens. This work investigates RL methods for optimizing sequential cancer chemotherapy decisions. A traditional Q‑learning based simulator is first constructed, discretizing drug dosages and administration times, and framing the problem as a Markov decision process with state‑action‑reward feedback.\n\nThe study then adopts an offline RL pipeline, employing Conservative Q‑Learning (CQL) and Model‑Based Offline Policy Optimization (MOPO) on real colon‑cancer clinical data. A learned transition model predicts patient blood‑work dynamics, while CQL adds a conservatism term to the target Q‑values to mitigate over‑estimation, and MOPO generates synthetic trajectories from the model to improve policy robustness. Experimental results indicate that the offline RL policies outperform the baseline Q‑learning in terms of survival outcomes and side‑effect management, offering a supportive decision‑making tool for junior oncologists.\n\nReview: The paper convincingly demonstrates the feasibility of offline reinforcement learning in high‑risk medical decision making. By combining conservative Q‑learning with model‑based policy optimization, it addresses data scarcity and safety constraints, paving a promising path toward personalized chemotherapy scheduling.