Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning
Hai Zhang, Boyuan Zheng, Tianying Ji, Jinhang Liu, Anqi Guo, Junqiao Zhao, Lanqing Li
TL;DR
The paper tackles generalization in offline meta-reinforcement learning (COMRL) by analyzing how context-based representations influence performance. It introduces task representation shift as a key factor that can break monotonic performance improvements and develops a return-discrepancy framework to connect context-encoder updates with the expected return $J^*(\theta)$. The authors derive a refined monotonic-improvement condition that explicitly accounts for drift in the task representation and propose a practical alternating framework with configurable update cadence ($N_{\text{k}},N_{\text{acc}}$) and encoder objectives (contrastive, reconstruction, cross-entropy) to rein in shift. Empirical results on MuJoCo and MetaWorld show performance gains when the task-representation drift is controlled, while analyses on pretraining and visualization caution against overinterpreting representations; the work lays groundwork for more robust COMRL training by explicitly modeling and managing representation shifts.
Abstract
Offline meta reinforcement learning (OMRL) has emerged as a promising approach for interaction avoidance and strong generalization performance by leveraging pre-collected data and meta-learning techniques. Previous context-based approaches predominantly rely on the intuition that alternating optimization between the context encoder and the policy can lead to performance improvements, as long as the context encoder follows the principle of maximizing the mutual information between the task variable $M$ and its latent representation $Z$ ($I(Z;M)$) while the policy adopts the standard offline reinforcement learning (RL) algorithms conditioning on the learned task representation.Despite promising results, the theoretical justification of performance improvements for such intuition remains underexplored.Inspired by the return discrepancy scheme in the model-based RL field, we find that the previous optimization framework can be linked with the general RL objective of maximizing the expected return, thereby explaining performance improvements. Furthermore, after scrutinizing this optimization framework, we observe that the condition for monotonic performance improvements does not consider the variation of the task representation. When these variations are considered, the previously established condition may no longer be sufficient to ensure monotonicity, thereby impairing the optimization process.We name this issue task representation shift and theoretically prove that the monotonic performance improvements can be guaranteed with appropriate context encoder updates.Our work opens up a new avenue for OMRL, leading to a better understanding between the task representation and performance improvements.