LAPO: Latent-VariableAdvantage-WeightedPolicy OptimizationforOfflineReinforcementLearning

But in practice, it requires querying the behavior policy which is unknown, and using an erroneous approximation of the behavior policy can negatively affect the performance ([39]).