Whenever you're ready, walk me through how you'd approach this.
How to improve the ETA model using reinforcement learning and RLHF?
Asked in the MLE system design stage. Discussed solutions and approaches for lowering loss in ETA prediction. Explain how you would determine whether reinforcement learning or RLHF is appropriate, define the objective and feedback signals, design offline training and online serving, and evaluate the resulting ETA model. Address data collection, delayed outcomes, safety constraints, and failure modes such as reward misspecification, feedback bias, feature drift, and training-serving skew.