Lessons on off-policy methods from a notification component of a chatbot

This work serves as a review of our experience applying off-policy techniques to train and evaluate a contextual bandit model powering a troubleshooting notification in a chatbot. First, we demonstrate the effectiveness of off-policy evaluation when data volume is orders of magnitude less than typically found in the literature. We present our reward function and choices behind its design, as well as how we construct our logging policy to balance exploration and performance on key metrics. Next, we present a guided framework to update a model post-training called Post-Hoc Reward Distribution Hacking, which we employed to improve model performance and correct deficiencies in trained models stemming from the existence of a null action and a noisy reward signal. Throughout the work, we include discussions of various practical pitfalls encountered while using off-policy methods in hopes to expedite other applications of these techniques.