Optimize Planning Heuristics to Rank, not to Estimate Cost-to-Goal
Authors: Leah Chrestien, Stefan Edelkamp, Antonin Komenda, Tomas Pevny
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental comparison on a diverse set of problems unequivocally supports the derived theory. |
| Researcher Affiliation | Collaboration | Leah Chrestien Czech Technical University in Prague leah.chrestien@aic.fel.cvut.cz Tomá s Pevný Czech Technical University in Prague, and Gen Digital, Inc. pevnytom@fel.cvut.cz Stefan Edelkamp Czech Technical University in Prague edelkste@fel.cvut.cz Antonín Komenda Czech Technical University in Prague antonin.komenda@fel.cvut.cz |
| Pseudocode | Yes | 2.1 Forward search algorithm: 1. Initialise Open list as O0 = {s0}. 2. Set g(s0) = 0 3. Initiate the Closed list to empty, i.e. C0 = . 4. For i 1, . . . until Oi = (a) Select the state si = arg mins Oi 1 αg(s) + βh(s) (b) Remove si from Oi 1, Oi = Oi 1 \ {si} (c) If si S , i.e. it is a goal state, go to 5. (d) Insert the state si to Ci 1, Ci = Ci 1 {si} (e) Expand the state si into states s for which hold (si, s ) E and for each i. set g(s ) = g(si) + w(si, s ) ii. if s is in the Closed list as sc and g(s ) < g(sc) then sc is reopened (i.e., moved from the Closed to the Open list), else continue with (e) iii. if s is in the Open list as so and g(s ) < g(so) then so is updated (i.e., removed from the Open list and re-added in the next step with updated g( )), else continue with (e) iv. add s into the Open list 5. Walk back to retrieve the solution path. |
| Open Source Code | Yes | The source code of all experiments is available at https://github.com/aicenter/Optimize-Planning-Heuristics-to-Rank with MIT license. |
| Open Datasets | Yes | The problem instances were generated by [40]. (refers to https://github.com/mp Schrader/gym-sokoban, 2018), The mazes were generated using an open-source maze generator [47] (refers to https://github.com/ravenkls/Maze-Generator-and-Solver, 2022), These puzzles were taken from [37]. (refers to https://github.com/levilelis/h-levin, 2021), The problem instances were copied from [44] (refers to https://github.com/williamshen-nz/STRIPS-HGN, 2021) (Blocksworld, N-Puzzle, and Ferry) and from [34] (refers to https://github.com/AI-Planning/ classical-domains, 2023) (elevators-00-strips). |
| Dataset Splits | Yes | The problem instances were randomly divided into training, validation, and testing sets, each containing 50% / 25% / 25% of the whole set of problems. |
| Hardware Specification | Yes | training used an NVIDIA Tesla GPU model V100-SXM2-32GB., All training and evaluation were done on single-core Intel Xeon Silver 4110 CPU 2.10GHz with a memory limit of 128GB. |
| Software Dependencies | Yes | The experiments were conducted in the Keras-2.4.3 framework with Tensorflow-2.3.1 as the backend. |
| Experiment Setup | Yes | The proposed ranking losses are instantiated with the logistic loss function as in Equation (4) for A* search by setting α = β = 1 in (2), called L , and for GBFS, by setting α = 0, β = 1 in (2), called Lgbfs., grid-search over the number of hyper-graph convolutions in {1, 2, 3}, dimensions of filters in {4, 8, 16}, ADAM [26] training algorithm was run for 100 20000 steps for the grid domains., Forward search algorithms were given 10 mins to solve each problem instance. |