A taxonomy of reinforcement learning faults
الملخص
Reinforcement Learning (RL) is increasingly being used for a variety of applications, some of which include safety-critical tasks. Therefore, it is important for developers to be aware of faults that occur when writing programs that leverage RL. This thesis conducts an empirical study on the faults associated with RL programs. Our work is tailored towards developers who use well-vetted frameworks or implementations of RL algorithms. We conduct studies to identify relevant RL tags to mine StackExchange and a popular RL framework to mine GitHub. Then, we mine StackExchange and GitHub and after dropping unrelated posts using various automated techniques, we arrive at 2,787 posts, which we manually analyze. Finally, we manually label a portion of these posts to develop a taxonomy of RL faults. Our final taxonomy contains 8 categories of faults from 112 artifacts. We describe these categories and the corresponding child nodes in detail in this thesis. We compare our work against prior work done in this area by comparing each leaf node of theirs with our taxonomy. We present two promising future directions that could be derived from this work: Mutation Testing and Fault Localization. Both of these directions require a taxonomy of real RL faults, which our work provides.
الكلمات المفتاحية
غير متوفر
التخصصات
رسائل وأطاريح مشابهة
- Reinforcement Learning in Industrial Applications Niklas Kotarsky ; Eric Bergvall · Lund University · 2020
- Reinforcement Learning on Resource Bounded Systems Jaden B. Travnik · University of Alberta · 2018
- Learning-based multi-robot collaboration for intelligent and adaptive robotic systems Yutong Du · Aalto University · 2025
- Securing Federated Learning: A Comprehensive Defence Against Privacy Attacks Moyeen, M A · Concordia University · 2025
- Toward Robot Learning for Fine Manipulation Jinda Cui · 2022
- Robot Learning for Muscular Systems Dieter Büchler · Technical University of Darmstadt · 2019
