pyastar2d 1.0.0

A simple implementation of the A* algorithm for path-finding on a two-dimensional grid.

PyAstar2D

This is a very simple C++ implementation of the A* algorithm for pathfinding on a two-dimensional grid. The solver itself is implemented in C++, but is callable from Python. This combines the speed of C++ with the convenience of Python.

I have not done any formal benchmarking, but the solver finds the solution to a 1802 by 1802 maze in 0.29s and a 4008 by 4008 maze in 0.83s when running on my nine-year-old Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz. See Example Results for more details.

See src/cpp/astar.cpp for the core C++ implementation of the A* shortest path search algorithm, src/pyastar2d/astar_wrapper.py for the Python wrapper and examples/example.py for example usage.

When determining legal moves, 4-connectivity is the default, but it is possible to set allow_diagonal=True for 8-connectivity.

Installation

If running on Linux or MacOS, simply run

pip install .

from the root directory. If you are using Windows you may have to install Cython manually first:

pip install Cython
pip install .

To check that everything worked, run the example:

python examples/example.py

You could also add it to your project by adding this line to requirements.txt:

pyastar2d @ git+git://github.com/hjweide/pyastar2d.git@master#egg=pyastar2d

Usage

A simple example is given below:

import numpy as np
import pyastar2d
# The minimum cost must be 1 for the heuristic to be valid.
# The weights array must have np.float32 dtype to be compatible with the C++ code.
weights = np.array([[1, 3, 3, 3, 3],
                    [2, 1, 3, 3, 3],
                    [2, 2, 1, 3, 3],
                    [2, 2, 2, 1, 3],
                    [2, 2, 2, 2, 1]], dtype=np.float32)
# The start and goal coordinates are in matrix coordinates (i, j).
path = pyastar2d.astar_path(weights, (0, 0), (4, 4), allow_diagonal=True)
print(path)
# The path is returned as a numpy array of (i, j) coordinates.
array([[0, 0],
       [1, 1],
       [2, 2],
       [3, 3],
       [4, 4]])

Note that all grid points are represented as (i, j) coordinates. An example of using pyastar2d to solve a maze is given in examples/maze_solver.py.

Example Results

To test the implementation, I grabbed two nasty mazes from Wikipedia. They are included in the mazes directory, but are originally from here: Small and Large. I load the .png files as grayscale images, and set the white pixels to 1 (open space) and the black pixels to INF (walls).

To run the examples specify the input and output files using the --input and --output flags. For example, the following commands will solve the small and large mazes:

python examples/maze_solver.py --input mazes/maze_small.png --output solns/maze_small.png
python examples/maze_solver.py --input mazes/maze_large.png --output solns/maze_large.png

Small Maze (1802 x 1802):

time python examples/maze_solver.py --input mazes/maze_small.png --output solns/maze_small.png
Loaded maze of shape (1802, 1802) from mazes/maze_small.png
Found path of length 10032 in 0.292794s
Plotting path to solns/maze_small.png
Done

real	0m1.214s
user	0m1.526s
sys	0m0.606s

The solution is visualized below:

Large Maze (4002 x 4002):

time python examples/maze_solver.py --input mazes/maze_large.png --output solns/maze_large.png
Loaded maze of shape (4002, 4002) from mazes/maze_large.png
Found path of length 783737 in 0.829181s
Plotting path to solns/maze_large.png
Done

real	0m29.385s
user	0m29.563s
sys	0m0.728s

The solution is visualized below:

Motivation

I recently needed an implementation of the A* algorithm in Python to find the shortest path between two points in a cost matrix representing an image. Normally I would simply use networkx, but for graphs with millions of nodes the overhead incurred to construct the graph can be expensive. Considering that I was only interested in graphs that may be represented as two-dimensional grids, I decided to implement it myself using this special structure of the graph to make various optimizations. Specifically, the graph is represented as a one-dimensional array because there is no need to store the neighbors. Additionally, the lookup tables for previously-explored nodes (their costs and paths) are also stored as one-dimensional arrays. The implication of this is that checking the lookup table can be done in O(1), at the cost of using O(n) memory. Alternatively, we could store only the nodes we traverse in a hash table to reduce the memory usage. Empirically I found that replacing the one-dimensional array with a hash table (std::unordered_map) was about five times slower.

Tests

The default installation does not include the dependencies necessary to run the tests. To install these, first run

pip install -r requirements-dev.txt

before running

py.test

The tests are fairly basic but cover some of the more common pitfalls. Pull requests for more extensive tests are welcome.

References

1. A* search algorithm on Wikipedia
2. Pathfinding with A* on Red Blob Games