Bounding-Box Inference for Error-Aware Model-Based Reinforcement Learning
This work addresses a critical bottleneck in model-based RL for agents that need to avoid performance degradation from model errors, though it is incremental as it builds on existing uncertainty estimation methods.
The paper tackles the problem of catastrophic interference in model-based reinforcement learning due to inaccurate models by proposing bounding-box inference for estimating model uncertainty, enabling selective planning that reliably improves performance.
In model-based reinforcement learning, simulated experiences from the learned model are often treated as equivalent to experience from the real environment. However, when the model is inaccurate, it can catastrophically interfere with policy learning. Alternatively, the agent might learn about the model's accuracy and selectively use it only when it can provide reliable predictions. We empirically explore model uncertainty measures for selective planning and show that best results require distribution insensitive inference to estimate the uncertainty over model-based updates. To that end, we propose and evaluate bounding-box inference, which operates on bounding-boxes around sets of possible states and other quantities. We find that bounding-box inference can reliably support effective selective planning.