When CEGAR Meets Regression: A Love Story in Optimal Classical Planning

Authors: Martín Pozo, Alvaro Torralba, Carlos Linares Lopez

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments We implemented our new refinement algorithms within the Scorpion planner (Seipp, Keller, and Helmert 2020). Our experiments run on the Autoscale 21.11 benchmark set (Torralba, Seipp, and Sievers 2021), which contains 42 domains with 30 tasks each, for a total of 1260 instances. All experiments are limited to 30 minutes and 8 GB of RAM and run in an Ubuntu Linux 20.04 server with an Intel Xeon X3470 processor at 2.93 GHz, 16 GB of RAM and a 1 TB HDD.
Researcher Affiliation Academia Mart ın Pozo1, Alvaro Torralba2, Carlos Linares L opez1 1Universidad Carlos III de Madrid, Madrid, Spain 2Aalborg University, Aalborg, Denmark
Pseudocode Yes Algorithm 1: CEGAR main algorithm
Open Source Code Yes Details are shown in the supplementary material, available in Zenodo along with code and experimental data (Pozo, Torralba, and Linares L opez 2023).
Open Datasets Yes Our experiments run on the Autoscale 21.11 benchmark set (Torralba, Seipp, and Sievers 2021), which contains 42 domains with 30 tasks each, for a total of 1260 instances.
Dataset Splits No The paper states that experiments run on the 'Autoscale 21.11 benchmark set', which contains 42 domains with 30 tasks each, but does not provide specific training/validation/test dataset splits.
Hardware Specification Yes All experiments are limited to 30 minutes and 8 GB of RAM and run in an Ubuntu Linux 20.04 server with an Intel Xeon X3470 processor at 2.93 GHz, 16 GB of RAM and a 1 TB HDD.
Software Dependencies No The paper mentions 'Ubuntu Linux 20.04 server' and the 'Scorpion planner', but does not provide specific version numbers for other key software components or libraries.
Experiment Setup Yes Unless explicitly said otherwise, we use Scorpion s default parameters for CEGAR. Specifically, we use incremental search for optimal abstract plans instead of A (Seipp, von Allmen, and Helmert 2020); set the termination condition as 1 million of non-looping transitions; and choose splits by maximizing the amount of covered flaws (COVER), breaking ties by the most refined split (MAX REFINED), i.e., the one with minimal remaining values/domain size (Speck and Seipp 2022).