Conference item
Verifying reinforcement learning up to infinity
- Abstract:
- Formally verifying that reinforcement learning systems act safely is increasingly important, but existing methods only verify over finite time. This is of limited use for dynamical systems that run indefinitely. We introduce the first method for verifying the time-unbounded safety of neural networks controlling dynamical systems. We develop a novel abstract interpretation method which, by constructing adaptable template-based polyhedra using MILP and interval arithmetic, yields sound - safe and invariant - overapproximations of the reach set. This provides stronger safety guarantees than previous time-bounded methods and shows whether the agent has generalised beyond the length of its training episodes. Our method supports ReLU activation functions and systems with linear, piecewise linear and non-linear dynamics defined with polynomial and transcendental functions. We demonstrate its efficacy on a range of benchmark control problems.
- Publication status:
- Published
- Peer review status:
- Peer reviewed
Actions
Access Document
- Files:
-
-
(Preview, Accepted manuscript, 1.1MB, Terms of use)
-
- Publisher copy:
- 10.24963/ijcai.2021/297
Authors
- Publisher:
- International Joint Conferences on Artificial Intelligence Organization
- Host title:
- Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI 2021)
- Journal:
- Proceedings of the International Joint Conference on Artificial Intelligence More from this journal
- Pages:
- 2154-2160
- Publication date:
- 2021-08-11
- Acceptance date:
- 2021-04-29
- Event title:
- 30th International Joint Conference on Artificial Intelligence (IJCAI 2021)
- Event location:
- Montreal, Canada
- Event website:
- https://ijcai-21.org/
- Event start date:
- 2021-08-21
- Event end date:
- 2021-08-26
- DOI:
- Language:
-
English
- Keywords:
- Pubs id:
-
1178085
- Local pid:
-
pubs:1178085
- Deposit date:
-
2021-05-23
Terms of use
- Copyright holder:
- IJCAI
- Copyright date:
- 2021
- Rights statement:
- © 2021, IJCAI
- Notes:
- This paper was presented at the 30th International Joint Conference on Artificial Intelligence (IJCAI-21), 21st - 26th August 2021, Montreal, Canada. This is the accepted manuscript version of the article. The final version is available from the conference proceedings at: https://doi.org/10.24963/ijcai.2021/297
If you are the owner of this record, you can report an update to it here: Report update to this record