Daniel Gnad

Abstract
Planning is a core challenge in Artificial Intelligence. Developing systems that can achieve complex goals autonomously requires the capability of planning and thinking ahead. While recent advances in large generative models have shown outstanding performance in many
applications, they still fail to solve complex planning problems. AI planning in the form of model-based reasoning has focused on enabling
such capabilities for many years, but, in practice, existing solutions often fail to scale to larger problems. By analyzing the inherent structure of the given problem, it is possible to push the limits of reasoning algorithms. This can take several forms that work on different levels of the algorithms. I will present two paradigms that exploit problem structure to make planning as state-space search more efficient. The first one is decoupled search, which decomposes the problem by analyzing the dependencies between model components. This leads to an exponential reduction in search effort, as it avoids enumerating reorderings of independent transitions. The second approach is exploiting problem structure to efficiently compute well-informed heuristics that guide the search process. This is done by identifying structurally simple components for which exact solutions can be computed in polynomial time, which leads to enhanced search performance and overall better scaling behavior.

Daniel Gnad is assistant professor in the Machine Reasoning Lab at Linköping University in Sweden. He did his studies in Computer Science
at Saarland University in Germany, where, after finishing his M.Sc. degree, he joined the Foundations of Artificial Intelligence group as a
Ph.D. student. His Ph.D. thesis won the ICAPS 2022 Best Dissertation Award and was nominated for the Dissertation Award 2021 of the German Society for Computer Science. Daniel's research interests are in the fields of AI planning and model checking with a common scheme of exploiting problem structure. He is working on state space reduction methods such as decoupled search, on model transformation techniques that enhance the problem specification, and structural analyses of classical and numeric planning formalisms. His planning systems won several awards at the recent International Planning Competitions.