Plan Recognition as Planning Revisited
Authors: Shirin Sohrabi, Anton V. Riabov, Octavian Udrea
IJCAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our two proposed plan recognition approaches, LPG-d [Nguyen et al., 2012] for diverse planning, and TK [Riabov et al., 2014] for top-k planning, against previous work [Ram ırez and Geffner, 2010]. We configured the previous work so that it uses the LM-Cut1 [Pommerening and Helmert, 2012] planner. We selected the LMCut planner because it was shown to perform well in the planning competition and it performs better than the original planners used for the configuration of the previous approach. Note, using a more efficient optimal planner or even choosing a sub-optimal planner will not improve the presented results for the previous approach because LM-Cut was able to solve all problems optimally within the time limit. We used a timeout of 30 minutes and ran all our experiments on a dual 16-core 2.70 GHz Intel(R) Xeon(R) E5-2680 processor with 256 GB RAM. |
| Researcher Affiliation | Industry | Shirin Sohrabi Anton V. Riabov Octavian Udrea IBM T.J. Watson Research Center 1101 Kitchawan Rd, Yorktown Heights, NY 10598, USA {ssohrab, riabov, udrea}@us.ibm.com |
| Pseudocode | No | The paper describes the transformation and definitions but does not provide structured pseudocode or an algorithm block. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code or links to a code repository for the described methodology. |
| Open Datasets | Yes | We used the Kitchen, Intrusion, Campus, and IPC-Grid domain from the previous approach [Ram ırez and Geffner, 2010] but modified the kitchen domain to disallow the extra take actions that are not towards achieving the goal in order to evaluate the approaches in the case of noisy observations that need to be discarded. |
| Dataset Splits | No | The paper describes generating problem instances by randomly selecting percentages of observations (25%, 50%, 75%) and adding noise, but it does not specify traditional training, validation, and testing dataset splits for machine learning models. |
| Hardware Specification | Yes | We used a timeout of 30 minutes and ran all our experiments on a dual 16-core 2.70 GHz Intel(R) Xeon(R) E5-2680 processor with 256 GB RAM. |
| Software Dependencies | No | The paper mentions specific planners used, such as 'LM-Cut1 [Pommerening and Helmert, 2012] planner', 'LPG-d [Nguyen et al., 2012]', and 'TK [Riabov et al., 2014]', but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For the LPG-d planner we used three settings of (m, d), (10, 0.75), (50, 0.5), and (100, 0.75), but report only on the (50, 0.5) case; 50 plans that are at least 0.5 distance away from each other. For the top-k planner, we set k = 1000 in order to obtain a large set of high-quality plans that are also diverse. Lower values of k might help improve the posterior probabilities but will likely result in lower converge. For the coefficients, we set b1 = 2, and b2 = 4 in the experiments meaning that we assign a higher penalty for the unexplained observations, and a lower penalty for the missing observations. |