29.11.2024
The talk was presented by Milad Kazemi from Department of Informatics at King's College London. He is an interdisciplinary researcher in Artificial Intelligence & Cyber-Physical Systems Research Associate. He made his BSc and MSc in Mechanical Engineering and a PhD in Computer Science. His research focuses on data-driven approaches for control synthesis in cyber-physical systems and explores counterfactual analysis in sequential decision-making. By integrating control theory, optimization, formal methods, causality, and machine learning, Milad’s work addresses key challenges in fields like medical devices, smart grids, and smart homes.
The talk explored limitations of existing counterfactual inference methods for policy evaluation in reinforcement learning, particularly in safety-critical domains. Existing approaches often rely on specific causal models, e.g., the Gumbel-max structural causal model (SCM), to ensure counterfactuals are identifiable. However, if these are not appropriate models of the underlying Markov Decision Process (MDP), this can lead to inaccurate counterfactual inference. We introduced a novel partial counterfactual inference approach that instead bounds counterfactual transition probabilities over all possible causal models. By incorporating assumptions like counterfactual stability and monotonicity, we generated more realistic probabilities and robust policies. We evaluate our approach using synthetic and real-world data, demonstrating improved robustness compared to existing methods. The presentation was based on ongoing research.
Thank you for the visit at RC Trust!