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 [1].
Maintenance of Plan Libraries for Case-Based Planning: Offline and Online Policies
Authors: Alfonso Emilio Gerevini, Alessandro Saetti, Ivan Serina, Andrea Loreggia, Luca Putelli, Anna Roubickova
JAIR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Experimental Results We present here a thorough collection of experiments. |
| Researcher Affiliation | Academia | Alfonso Emilio Gerevini EMAIL Alessandro Saetti EMAIL Ivan Serina EMAIL Andrea Loreggia EMAIL Luca Putelli EMAIL Universit a degli Studi di Brescia Brescia, Italy Anna Roub ıˇckov a EMAIL The University of Edinburgh Edinburgh, The United Kingdom |
| Pseudocode | Yes | Figure 3: Schema of a greedy algorithm computing an approximation of the Coverage-Based Policy. Figure 5: Algorithm for the bounded online maintenance policy. |
| Open Source Code | No | The problems and solutions used to construct the libraries, the libraries used for the experimental study, the test problems, as well as the executable of OAKplan, are made available from https://lpg.unibs.it/OAKplan/. |
| Open Datasets | Yes | The benchmark domains used in our experimental analysis are Driverlog, Logistics, Rovers, Satellite, Zenotravel and Elevators. The first five domains are from the 3rd international planning competition... The sixth domain, Elevators, is from the 6th international planning competition (Helmert, Do, & Refanidis, 2008)... The problems and solutions used to construct the libraries, the libraries used for the experimental study, the test problems, as well as the executable of OAKplan, are made available from https://lpg.unibs.it/OAKplan/. |
| Dataset Splits | Yes | for each considered domain we generated two sets of 6000 cases each, L1 and L2. The cases in L1 are grouped into a high number of small and middle-size clusters, while the cases in L2 are grouped into a much lower number of middle-size and large clusters. ... Moreover, for each considered domain, we generated 25 test problems obtained from randomly selected problems in the libraries by (i) changing all object names, and (ii) performing a number of modifications to the problem propositions (initial facts and/or goals) ranging from 1 to 10. |
| Hardware Specification | Yes | All the experimental tests were run on an Intel Xeon(tm) 2 GHz machine with 20 Gbytes of RAM. |
| Software Dependencies | No | The techniques presented in the previous sections have been implemented in a new version of the CBP system OAKplan (Serina, 2010) that, to the best of our knowledge, is the stateof-the-art for CBP systems. In our experiments, the plan retrieved by OAKplan is adapted using planner LPG-td (Gerevini, Saetti, & Serina, 2003, 2006, 2011; Fox, Gerevini, Long, & Serina, 2006)... The plan libraries of these domains were generated using planner TLPlan (Bacchus & Kabanza, 2000)... the plans in the library for Elevators were generated by FF (Hoffmann, 2003)... |
| Experiment Setup | Yes | Specifically, a given case base of 6000 cases is reduced to libraries of 3000, 2000, 1000, 500, 250, and 100 cases using values of δ estimated as the 50th, 67th, 83th, 92th, 96th, and 98th percentiles of the average minimum distance distribution, respectively. The CPU-time limit for each run of the experimented planners was 30 minutes, after which termination was forced. |