Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Optimizing the Optimization of Planning Domains by Automatic Action Schema Splitting
Authors: Mojtaba Elahi, Jussi Rintanen
AAAI 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted two experiments to evaluate the effectiveness of our approach 2. In the first experiment, we sought to evaluate the contribution of each of our proposed components in solving hard-to-ground problems. In the second experiment, our goal was to understand how our approach influences the performance of planners when dealing with a broader range of problems. We chose four planners with different technologies for our experiments: LAMA (Richter and Westphal 2010), a heuristic-based planner; Maidu (Corrˆea et al. 2023b), a portfolio-based planner and the winner of IPC 23 in the classical-satisficing track; a variant of BFWS (Lipovetzky and Geffner 2017), featuring width-based search; and Powerlifted (Corrˆea et al. 2023a), the state-of-the-art planner for hard-to-ground problems. The variant of BFWS that we used in our experiment is a combination of two modes of this planner. First, the planner is run in goalcount-only mode for at most 15 minutes, then it switches to f5-initstate-relevant. Additionally, we used a slightly modified version of this planner with more efficient memory usage. In our experiments, we ran the planners with a 30 minute time limit on 2.50 GHz Intel Processor with an 8 GB memory limit. To determine the threshold ω in our approach, we roughly estimated that the upper bound for the number of ground actions that a planner can deal with is 106. Therefore, we set ω = 106 |A| , where A is the set of action schemas. Also, we considered a time limit of max(50, 500 |A| ) for the Monte Carlo search. In the report of our experimental results, we refer to our approach as EAS (Enhanced Action Splitter). |
| Researcher Affiliation | Academia | Mojtaba Elahi, Jussi Rintanen Aalto University EMAIL, EMAIL |
| Pseudocode | No | The paper describes algorithms and methods in prose but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at: https://github.com/melahi/enhanced-action-splitter |
| Open Datasets | Yes | We conducted our experiment on a set of hard-to-ground domains, including Pipesworld-Tankage, Genome Edit Distance (GED), Organic Synthesis-Alkene, Organic Synthesis-MIT, and Organic Synthesis-Original. [...] with a large set of domains, including hard-to-ground domains and domains from IPC 14, IPC 18, and IPC 23 |
| Dataset Splits | No | The paper mentions |
| Hardware Specification | Yes | In our experiments, we ran the planners with a 30 minute time limit on 2.50 GHz Intel Processor with an 8 GB memory limit. |
| Software Dependencies | No | The paper does not list specific version numbers for software dependencies used in the experiments (e.g., Python, specific libraries or frameworks). |
| Experiment Setup | Yes | In our experiments, we ran the planners with a 30 minute time limit on 2.50 GHz Intel Processor with an 8 GB memory limit. To determine the threshold ω in our approach, we roughly estimated that the upper bound for the number of ground actions that a planner can deal with is 106. Therefore, we set ω = 106 |A| , where A is the set of action schemas. Also, we considered a time limit of max(50, 500 |A| ) for the Monte Carlo search. |