Safe Reinforcement Learning via Probabilistic Logic Shields
2023 (English)In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI 2023) / [ed] Edith Elkind, International Joint Conferences on Artificial Intelligence , 2023, p. 5739-5749Conference paper, Published paper (Refereed)
Abstract [en]
Safe Reinforcement learning (Safe RL) aims at learning optimal policies while staying safe. A popular solution to Safe RL is shielding, which uses a logical safety specification to prevent an RL agent from taking unsafe actions. However, traditional shielding techniques are difficult to integrate with continuous, end-to-end deep RL methods. To this end, we introduce Probabilistic Logic Policy Gradient (PLPG). PLPG is a model-based Safe RL technique that uses probabilistic logic programming to model logical safety constraints as differentiable functions. Therefore, PLPG can be seamlessly applied to any policy gradient algorithm while still providing the same convergence guarantees. In our experiments, we show that PLPG learns safer and more rewarding policies compared to other state-of-the-art shielding techniques.
Place, publisher, year, edition, pages
International Joint Conferences on Artificial Intelligence , 2023. p. 5739-5749
Keywords [en]
Uncertainty in AI, UAI, Statistical relational AI Knowledge Representation and Reasoning, KRR, Learning and reasoning Machine Learning, ML, Reinforcement learning
National Category
Computer Sciences
Research subject
Computer Science
Identifiers
URN: urn:nbn:se:oru:diva-111102DOI: 10.24963/ijcai.2023/637ISI: 001202344205093Scopus ID: 2-s2.0-85170386139ISBN: 9781956792034 (electronic)OAI: oai:DiVA.org:oru-111102DiVA, id: diva2:1831952
Conference
32nd International Joint Conference on Artificial Intelligence (IJCAI 2023), Macao, S.A.R., August 19-25, 2023
Funder
Wallenberg AI, Autonomous Systems and Software Program (WASP)EU, Horizon 2020, 952215
Note
This work was supported by the Research Foundation - Flanders under EOS No. 309925744, the Flemish Government (AI Research Program), the EU Horizon 2020 programme TAILOR under No. 952215, and the KU Leuven Research fund. GM has received funding from FWO (1239422N). LDR is partially funded by the Wallenberg AI, Autonomous Systems and Software Program.
2024-01-272024-01-272025-01-16Bibliographically approved