2021/11/09 by Stefan Radic Webster, Webster, Stefan Radic, Peter Flach +1
Computer Science · Decision Sciences · #Adversarial Robustness in Machine Learning #Complex Systems and Decision Making #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Reinforcement Learning in Robotics
paper · pdf · doi:10.48550/arxiv.2111.04972
openalex publication_date 2021/11/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Identifying uncertainty and taking mitigating actions is crucial for safe and\ntrustworthy reinforcement learning agents, especially when deployed in\nhigh-risk environments. In this paper, risk sensitivity is promoted in a\nmodel-based reinforcement learning algorithm by exploiting the ability of a\nbootstrap ensemble of dynamics models to estimate environment epistemic\nuncertainty. We propose uncertainty guided cross-entropy method planning, which\npenalises action sequences that result in high variance state predictions\nduring model rollouts, guiding the agent to known areas of the state space with\nlow uncertainty. Experiments display the ability for the agent to identify\nuncertain regions of the state space during planning and to take actions that\nmaintain the agent within high confidence areas, without the requirement of\nexplicit constraints. The result is a reduction in the performance in terms of\nattaining reward, displaying a trade-off between risk and return.\n