This is a codebase for collecting non-parametric heuristic data error model. We run suboptimal search algorithm such as weighted A* and Lsslrta* to collect observed states. Then use optimal solver such as A*, IDA*(CR) (sometime with stronger heuristc) to compute the optimal cost-to-go value. In short, we generate belief distribution of h* values for h.
[1] Maximilian Fickert, Tianyi Gu, Leonhard Staut, Wheeler Ruml, Joerg Hoffmann, and Marek Petrik, Beliefs We Can Believe In: Replacing Assumptions with Data in Real-Time Search. Proceedings of the Thirty-fourth AAAI Conference on Artificial Intelligence (AAAI-20), 2020.
[2] Maximilian Fickert, Tianyi Gu, Leonhard Staut, Sai Lekyang, Wheeler Ruml, Joerg Hoffmann, and Marek Petrik, Real-time Planning as Data-driven Decision-making. Proceedings of the ICAPS Workshop on Bridging the Gap Between AI Planning and Reinforcement Learning (PRL-20), 2020.
[pdf] [talk] [slides] [poster]
mkdir ../realtime-nancy
cd ../realtime-nancy
mv ../nancybackup .
mkdir build_release
mkdir build_debug
mkdir plots
mkdir results
mkdir worlds
make-
generate domain instances and put them into worlds folder
- tiles instances can be found here
- feel free to email me if you need other domain's instance or generator
-
run weighted A* or lsslrtA* on 500 random to gather all visited states
- single run:
cd ../build_release ./distributionPractice -o out.txt < ../world/slidingTile/1-4x4.st
- run batch of experiments:
cd script/testHarnesses/ ./singleThread-suboptimalSolver.sh -h
- single run:
-
check all observed states:
cd ../result/SlidingTilePuzzle/distributionTest ls -la
-
for each h value, collect 200 unique states according to their frequency counter
cd ../build_release ./collect -h
- check all collected problem
cd ../result/SlidingTilePuzzle/sampleProblem ls -la
- check all collected problem
-
use optimal solver to solve all the collected instances.
- For uniform puzzle, use idastar with manhattan heuristic
- For heavy and inverse puzzle, use ida star with pattern database heuristic
- For pancake and racetrack, use wastar with weight 1
on all ai machines:
cd script/testHarnessesheavy tile or inverse tile:
./multiThread-pdb.sh 2 1 67199 idastar heavy lsslrtastaruniform tile, pancake, or racetrack:
./multThread-optimalSolver.sh -hcheck all solutions
cd ../result/SlidingPuzzle/sampleData- compute h-h* distribution, and their sample sets
cd scripts/distributionGenerator/ ./distGenerator.py -h
This should generate a json file that contains the data that represent the histogram.
-
(optional) generate the post-search samples by doing one step lookahead on the samples and nancy backing-up their distribution and merging them.
cd ../build_release ./collect-post-search heavy lsslrtastar
-
(optional) compute post-search h-h* distribution based on samples merged by previous step
cd script/plotting python distribtuion.py uniform postd
This should generate another json file that contains the data that represent the histogram.
All the instances can be found here