Deep SPI: Safe Policy Improvement via World Models
This work addresses safe policy improvement for online reinforcement learning practitioners, offering a principled method with theoretical backing, though it builds incrementally on existing offline RL concepts.
The paper tackled the problem of safe policy improvement in online reinforcement learning by developing a theoretical framework and algorithm that ensures monotonic improvement and convergence. The result was DeepSPI, which matched or exceeded baselines like PPO and DeepMDPs on the ALE-57 benchmark while retaining theoretical guarantees.
Safe policy improvement (SPI) offers theoretical control over policy updates, yet existing guarantees largely concern offline, tabular reinforcement learning (RL). We study SPI in general online settings, when combined with world model and representation learning. We develop a theoretical framework showing that restricting policy updates to a well-defined neighborhood of the current policy ensures monotonic improvement and convergence. This analysis links transition and reward prediction losses to representation quality, yielding online, "deep" analogues of classical SPI theorems from the offline RL literature. Building on these results, we introduce DeepSPI, a principled on-policy algorithm that couples local transition and reward losses with regularised policy updates. On the ALE-57 benchmark, DeepSPI matches or exceeds strong baselines, including PPO and DeepMDPs, while retaining theoretical guarantees.