FMAI at EUMAS 2026: tutorial and paper accepted
Our tutorial "Formal Methods for Safe Reinforcement Learning" is part of the EASSS 2026 pre-conference programme at EUMAS 2026, and our paper on contract-based compositional shielding was also accepted to EUMAS 2026.
We are pleased to share two pieces of news from EUMAS 2026 in Malmö, Sweden: our tutorial Formal Methods for Safe Reinforcement Learning was accepted to EASSS 2026, the European Agent Systems Summer School held as part of the EUMAS pre-conference programme, and our paper Contract-Based Compositional Shielding for Safe Multi-Agent Reinforcement Learning was accepted to the 23rd European Conference on Multi-Agent Systems.
EASSS 2026 tutorial: Formal Methods for Safe Reinforcement Learning
Presenters: Dr. Francesco Belardinelli and Omar Adalat.
The tutorial presents the connection between reinforcement learning and formal methods, including constrained reinforcement learning, Linear Temporal Logic, automata-based specification, model checking, synthesis, and shielding. It focuses on how these tools can be used to design learning agents with formal objectives, constraints, and safety guarantees. Since the tutorial targets a school on multi-agent systems, a section is dedicated solely to multi-agent safety specification and shielding.
Tutorial slides | EASSS 2026 programme
Paper accepted at EUMAS 2026
Contract-Based Compositional Shielding for Safe Multi-Agent Reinforcement Learning
Authors: Omar Adalat, Edwin Hamel-De le Court, and Francesco Belardinelli.
The paper studies decentralised shielding for safe multi-agent reinforcement learning. It uses contract-based local safety obligations to enforce a shared global temporal-logic specification. We prove that, under mild representation conditions which preserve the optimal policy, the optimal team policy can be recovered in an end-to-end safe learning loop.