2025/12/19 by Schmöcker, Robin, Dockhorn, Alexander, Rosenhahn, Bodo
#004 #621.3 #Artificial intelligence #benchmark #search #sequential decision-making
paper · doi:10.15488/20257
Probabilistic planning research, such as improvements to Monte Carlo Tree Search, is faced with two challenges. Firstly, one usually requires an algorithm’s evaluation on a variety of test problems, trying several hundred or even thousands of parameter combinations, which in turn requires a large number of models that can be efficiently queried. Secondly, one has to identify a suitable set of problems to do the evaluation on. This work is made up of two components that together provide a solution for both problems for the probabilistic planning community. The first component of our work is software, which provides numerous models that are implemented by hand with domain-specific optimizations for optimal runtime. In particular, the software component contains 22 parametrized, non-deterministic Markov Decision Processes (MDP), which is currently the largest open-source, non-interpreter-based stochastic MDP library. The implementation provides access to methods required for planning (e.g., copying a state) as well as for reinforcement learning (e.g., getting an observation) through a C++ and a Python interface, which itself satisfies the Gymnasium interface. The second component of this work contains a detailed description as well as a statistical analysis of each model’s properties, which acts as a catalogization of the environments, enabling researchers to quickly select a set of environments satisfying their specific criteria.