[Re] FairDICE: A Gap Between Theory And Practice

A team of four researchers (Adema, Galliamov, Evstratovskiy, and Geurts) published a formal replication study of the FairDICE algorithm, a proposed method for achieving fairness in multi-objective offline reinforcement learning (RL). Their investigation, titled '[Re] FairDICE: A Gap Between Theory And Practice,' reveals a significant disconnect between the algorithm's theoretical promise and its initial practical implementation. While they confirm the theoretical claims hold, they discovered a critical error in the original codebase that fundamentally altered the method's behavior in continuous environments, reducing it to simple behavior cloning—a much less sophisticated technique. This bug meant the published experimental results did not accurately demonstrate FairDICE's intended capability to automatically learn fair trade-offs between competing objectives.

After rectifying the coding error and addressing underspecified hyperparameters, the replication team conducted extended experiments. They demonstrated that a corrected FairDICE can indeed scale to more complex environments and high-dimensional reward spaces, validating its underlying theoretical potential. However, their work highlights a heavy reliance on manual, often online, hyperparameter tuning to achieve these results, which poses practical deployment challenges. The study concludes that while FairDICE remains a theoretically interesting contribution to offline RL, its initial experimental justification requires substantial revision, underscoring the critical importance of rigorous code validation and detailed reporting in machine learning research to ensure replicability and trust in published claims.