Selecting effective feedback for learners at each stage of the programming problem-solving process remains an underexplored challenge in programming education, despite recent advances in automated feedback generation. This paper investigates feedback policy optimization using interaction logs collected from a deployed rule-based feedback system (over 25,000 interactions). We introduce an offline reinforcement learning approach that encodes distinct instructional objectives into reward functions. By modelling feedback selection as a Markov Decision Process (MDP) and training policies with Conservative Q-Learning (CQL), which enables robust learning from fixed datasets, we show that the learned policies yield consistently higher expected returns than the baseline rule-based policy across all evaluated objectives. Further analysis reveals that optimal feedback strategies vary systematically according to instructional objectives, underscoring the limitations of relying on a single, fixed feedback choice and the need for adaptive feedback selection that is systematically aligned with specified instructional objectives.