Towards Topologically Diverse Probabilistic Planning Benchmarks

Abstract

Markov Decision Processes (MDPs) are often used in Artificial Intelligence to solve probabilistic sequential decision-making problems. In the last decades, many probabilistic planning algorithms have been developed to solve MDPs. However, the lack of standardized benchmarks makes it difficult to compare the performance of these algorithms in different contexts. In this paper, we identify important topological properties of MDPs that can make a significant impact on the relative performance of probabilistic planning algorithms. We also propose a new approach to generate synthetic MDP domains having different topological properties. This approach relies on the connection between MDPs and graphs and allows every graph generation technique to be used to generate synthetic MDP domains.

Jaël Champagne Gareau

PhD Student in Computer Science

My research interests include AI, data structures, algorithms and HPC.