STAI CDT student Alexander Goodall (2023 cohort) presented the paper “Approximate Model-based Shielding for Safe Reinforcement Learning” at the 26th European Conference on Artificial Intelligence (ECAI 2023) in Kraków, Poland.
Alexander explains the paper here; “We developed a novel framework for ‘shielding’ reinforcement learning agents by ‘imagining’ possible future scenarios and overriding the agent with a corresponding backup or safe policy, if the probability of a bad event in the near future is too high. Our method comes with safety guarantees, meaning we exhibit safe behavior with high probability. The applications of this work are autonomous driving, robotics and industrial control systems.”
The conference provided opportunities to network, hear about exciting research and explore a new city. Alexander said, “The conference was great, there were a lot of interesting talks and I met a lot of fun people. The highlight has got to be the gala dinner, which was 90 meters underground in a salt mine! A very cool setting, especially for dinner. If you are ever in Krakow I would highly recommend going if only for a tour.”
You can read the full paper here: Alexander W Goodall; Francesco Belardinelli; Approximate Model-Based Shielding for Safe Reinforcement Learning.