Investigating the chaotic dynamics produced by deep reinforcement learning controllers
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Reinforcement Learning, Chaos Theory, Dynamical Systems, Robotics
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Deep reinforcement learning has had relatively few applications to real world problems as it lacks stability and performance guarantees. We address this issue and show the dynamics produced are deterministically chaotic.
Abstract: In recent years, deep Reinforcement Learning (RL) has demonstrated remarkable performance in simulated control tasks however there have been significantly fewer applications to real-world problems. While there are several reasons for this dichotomy, one key limitation is a need for theoretical stability guarantees in real-world applications, a property which cannot be provided by Deep Neural Network controllers. In this work, we investigate the stability of trained RL policies for continuous control tasks and identify the types of dynamics produced by the Markov Decision Process (MDP). We find the solutions produced by this interaction are deterministically chaotic with small initial inaccuracies in sensor readings or actuator movements compounding over time producing significantly different long-term outcomes, despite intervention in intermediate steps. The presence of these chaotic dynamics in the MDP provides evidence that RL controllers produce unstable solutions, limiting their application to real-world problems.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7715
Loading