This repository contains code that we used for experimenting with using artificial neural networks as a heuristic for solving the 15 puzzle, which we presented in the following scientific papers:
- On the Design of a Heuristic based on Artificial Neural Networks for the Near Optimal Solving of the (N2-1)-puzzle
- Near Optimal Solving of the (N2-1)-puzzle Using Heuristics Based on Artificial Neural Networks
Using the code in this repository we have collected a dataset (link) of optimal solution lengths to 6 million instances of 15 puzzle obtained via random permutations. Here are the first lines of the dataset (CSV format):
0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,cost
10,1,5,4,14,11,15,7,0,9,3,2,6,8,12,13,48
1,7,12,2,5,0,6,3,9,8,14,4,13,11,10,15,26
6,9,4,1,13,0,12,10,7,5,11,14,3,15,8,2,54
15,5,0,6,9,4,8,11,10,13,2,1,12,3,14,7,54
6,7,12,4,0,15,11,3,13,10,1,5,9,14,8,2,53
...
(6M total instances)
The first row is the header row. Pebbles are numbered from 1 to 15, pebble 0 represents an empty position. Positions are indexed from left to right, top to bottom, starting at 0. The first column, 0, represents the number of the pebble at position 0. An equivalent rule holds for columns 1 to 15. The final column, cost, represents the optimal (lowest) number of moves in which the puzzle instance can be solved.
For example, the row:
10,1,5,4,14,11,15,7,0,9,3,2,6,8,12,13,48
means that the instance
+----+----+----+----+
| 10 | 1 | 5 | 4 |
+----+----+----+----+
| 14 | 11 | 15 | 7 |
+----+----+----+----+
| | 9 | 3 | 2 |
+----+----+----+----+
| 6 | 8 | 12 | 13 |
+----+----+----+----+
can optimally be solved in 48 moves.
The row:
1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,0,0
would represent the solved puzzle:
+----+----+----+----+
| 1 | 2 | 3 | 4 |
+----+----+----+----+
| 5 | 6 | 7 | 8 |
+----+----+----+----+
| 9 | 10 | 11 | 12 |
+----+----+----+----+
| 13 | 14 | 15 | |
+----+----+----+----+
The dataset was collected using IDA* with the 7-8 pattern database heuristic. The CSV (235 MB) can be downloaded from here.
- prototype - Python code with solver that uses the ANN-distance heuristic, and additional Jupyter notebooks with experiments
- evaluation - C++ app for compute-efficient implementations of algorithms (currently only used for generating the optimal cost datasets)
- utils - Utilities for working with datasets
- data - Directory for storing data (must be created manually)
- Set the current
PROJECT_ROOTin prototype/constants.py - Create the conda environment using the environment-cpu.yml or environment-gpu.yml file (use the latter if you have a CUDA compatible GPU)
- Activate the conda environment
- Choose a module and run it with python -m MODULE_NAME (for example, python -m prototype.experiments.random_boards_distribution)
- Enter the evaluation directory:
cd evaluation - Create a build directory and enter it:
mkdir build && cd build - Run cmake (with the project root directory as argument):
cmake .. - Build the project:
make - Run the program:
./fifteen_puzzle_solver(this currently only runs the dataset generator, which requires additional pattern database files TODO)
In case you find this repository or dataset helpful, feel free to cite our related publication Near Optimal Solving of the (N2-1)-puzzle Using Heuristics Based on Artificial Neural Networks:
@inbook{Cahlik2021,
title = {Near Optimal Solving of the (N2-1)-puzzle Using Heuristics Based on Artificial Neural Networks},
author = {Cahlik, Vojtech and Surynek, Pavel},
year = 2021,
booktitle = {Computational Intelligence: 11th International Joint Conference, IJCCI 2019, Vienna, Austria, September 17--19, 2019, Revised Selected Papers},
publisher = {Springer International Publishing},
address = {Cham},
pages = {291--312},
doi = {10.1007/978-3-030-70594-7_12},
isbn = {978-3-030-70594-7},
url = {https://doi.org/10.1007/978-3-030-70594-7_12}
}