Offline RL by Reward-Weighted Fine-Tuning for Conversation Optimization