Effort-Aware Benchmark Dataset for Sokoban Pathfinding Algorithms

This repository provides a comprehensive collection of datasets, pretrained models, solver outputs, and supporting resources focused on the effort-aware evaluation of classical pathfinding algorithms within procedurally generated Sokoban environments. The datasets capture solver behavior across six dimensions of computational effort — solution steps, cumulative path cost, number of node expansions, number of edge traversals, wall-clock runtime, and peak memory consumption.Evaluated solvers include Breadth-First Search (BFS), Depth-First Search (DFS), A* Search, and Greedy Best-First Search (GBFS), covering a broad spectrum of uninformed and heuristic search strategies. The repository contents include: Datasets: CSV files recording solver metrics across Sokoban levels of varying grid sizes (5×5, 6×6, 7×7, and 8×8), including merged datasets and difficulty-labeled datasets using clustering-based classification. Pretrained Models: KMeans clustering model trained on normalized effort metrics, uncovering structural patterns among levels. Min-Max feature scaler used during dimensionality reduction and model training. Solver-to-cluster mapping artifact supporting adaptive solver selection strategies. Meta-Solver Predictions: Prediction outputs associating early search features with the solver most likely to succeed on a given level. Documentation: Detailed README file explaining data schema, model usage, and instructions for reproducing key analyses and visualizations. Key experimental steps supported by this repository include effort-based clustering of problem instances, evaluation of solver performance trade-offs, identification of structural difficulty patterns in Sokoban levels, and prediction of solver suitability using lightweight effort-aware features. This repository supports reproducible research in effort-aware pathfinding, instance-aware algorithm selection, and adaptive planning strategies.It is intended to facilitate future studies in game AI, robotic motion planning, combinatorial search, and hybrid symbolic-learning planning frameworks. How to Cite This WorkIf you use this dataset, models, or any associated resources in your research, please cite this Zenodo record using the following DOI:10.5281/zenodo.15297570