Long-Horizon Model-Based Offline Reinforcement Learning Without Conservatism

摘要

Popular offline reinforcement learning (RL) methods rely on conservatism, penalizing out-of-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective. By modeling a posterior over plausible world models and training a history-dependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty and enables test-time generalization, without conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that long-horizon rollouts are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding errors, yielding our algorithm, NEUBAY, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-the-art on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods.