6 Jul 2014

Table of Contents

This article is a companion guide to my introduction to A*. I show how to implement Breadth-First Search, Dijkstra’s Algorithm, Greedy Best-First Search, and A*. I focused on keeping things simple.

#Python Implementation

I explain most of the code below. There are a few extra bits that you can find in implementation.py. These use Python 3 so if you use Python 2, you will need to change the super() call and the print function to the Python 2 equivalents.

1.1 Breadth First Search

Let’s implement Breadth First Search in Python. The main article shows the Python code for the search algorithm, but we also need to define the graph it works on. These are the abstractions I’ll use:

Graph
a data structure that can tell me the neighbors for each location (see this tutorial)
Locations
a simple value (int, string, tuple, etc.) that labels locations in the graph
Search
an algorithm that takes a graph, a starting location, and optionally a goal location, and calculates some useful information (visited, parent pointer, distance) for some or all locations
Queue
a data structure used by the search algorithm to decide the order in which to process the locations

In the main article, I focused on search. On this page, I’ll fill in the rest of the details to make complete working programs. Let’s start with a graph:

class SimpleGraph:
    def __init__(self):
        self.edges = {}
    
    def neighbors(self, id):
        return self.edges[id]

Yep, that’s all we need! You may be asking, where’s the Node object? The answer is: I rarely use a node object. I find it simpler to use integers, strings, or tuples as locations, and then use arrays or hash tables that use locations as an index.

Note that the edges are directed: we can have an edge from A to B without also having an edge from B to A. In game maps most edges are bidirectional but sometimes there are one-way doors or jumps off cliffs that are expressed as directed edges. Let’s make an example graph, where the locations are letters A-E.

implementation-example-graph.png

For each location I need a list of which locations it leads to:

example_graph = SimpleGraph()
example_graph.edges = {
    'A': ['B'],
    'B': ['A', 'C', 'D'],
    'C': ['A'],
    'D': ['E', 'A'],
    'E': ['B']
}

Before we can use it with a search algorithm, we need to make a queue:

import collections

class Queue:
    def __init__(self):
        self.elements = collections.deque()
    
    def empty(self):
        return len(self.elements) == 0
    
    def put(self, x):
        self.elements.append(x)
    
    def get(self):
        return self.elements.popleft()

This queue class is just a wrapper around the built-in collections.deque class. Feel free to use deque directly in your own code.

Let’s try the example graph with this queue and the breadth-first search algorithm code from the main article:

from implementation import *

def breadth_first_search_1(graph, start):
    # print out what we find
    frontier = Queue()
    frontier.put(start)
    visited = {}
    visited[start] = True
    
    while not frontier.empty():
        current = frontier.get()
        print("Visiting %r" % current)
        for next in graph.neighbors(current):
            if next not in visited:
                frontier.put(next)
                visited[next] = True

breadth_first_search_1(example_graph, 'A')

Grids can be expressed as graphs too. I’ll now define a new graph called SquareGrid, with locations tuples (int, int). Instead of storing the edges explicitly, I’ll calculate them in the neighbors function.

class SquareGrid:
    def __init__(self, width, height):
        self.width = width
        self.height = height
        self.walls = []
    
    def in_bounds(self, id):
        (x, y) = id
        return 0 <= x < self.width and 0 <= y < self.height
    
    def passable(self, id):
        return id not in self.walls
    
    def neighbors(self, id):
        (x, y) = id
        results = [(x+1, y), (x, y-1), (x-1, y), (x, y+1)]
        if (x + y) % 2 == 0: results.reverse() # aesthetics
        results = filter(self.in_bounds, results)
        results = filter(self.passable, results)
        return results

Let’s try it out with the first grid in the main article:

from implementation import *
g = SquareGrid(30, 15)
g.walls = DIAGRAM1_WALLS # long list, [(21, 0), (21, 2), ...]
draw_grid(g)

In order to reconstruct paths we need to store the location of where we came from, so I’ve renamed visited (True/False) to came_from (location):

from implementation import *

def breadth_first_search_2(graph, start):
    # return "came_from"
    frontier = Queue()
    frontier.put(start)
    came_from = {}
    came_from[start] = None
    
    while not frontier.empty():
        current = frontier.get()
        for next in graph.neighbors(current):
            if next not in came_from:
                frontier.put(next)
                came_from[next] = current
    
    return came_from

g = SquareGrid(30, 15)
g.walls = DIAGRAM1_WALLS

parents = breadth_first_search_2(g, (8, 7))
draw_grid(g, width=2, point_to=parents, start=(8, 7))

Some implementations use internal storage, creating a Node object to hold came_from and other values for each graph node. I’ve instead chosen to use external storage, creating a single hash table to store the came_from for all graph nodes. If you know your map locations have integer indices, another option is to use an array to store came_from.

1.2 Early Exit

Following the code from the main article, all we need to do is add an if statement in the main loop. This test is optional for Breadth First Search or Dijkstra’s Algorithm and effectively required for Greedy Best-First Search and A*:

from implementation import *

def breadth_first_search_3(graph, start, goal):
    frontier = Queue()
    frontier.put(start)
    came_from = {}
    came_from[start] = None
    
    while not frontier.empty():
        current = frontier.get()
        
        if current == goal:
            break
        
        for next in graph.neighbors(current):
            if next not in came_from:
                frontier.put(next)
                came_from[next] = current
    
    return came_from

g = SquareGrid(30, 15)
g.walls = DIAGRAM1_WALLS

parents = breadth_first_search_3(g, (8, 7), (17, 2))
draw_grid(g, width=2, point_to=parents, start=(8, 7), goal=(17, 2))

You can see that the algorithm stops when it finds the goal Z.

1.3 Dijkstra’s Algorithm

This is what adds complexity to graph search, because we’re going to start processing locations in a better order than “first in, first out”. What do we need to change?

  1. The graph needs to know cost of movement.
  2. The queue needs to return nodes in a different order.
  3. The search needs to keep track of these costs from the graph and give them to the queue.

1.3.1 Graph with weights

I’m going to add a cost(from_node, to_node) function that tells us the cost of moving from location from_node to its neighbor to_node. In this forest map I chose to make movement depend only on to_node, but there are other types of movement that use both nodes. An alternate implementation would be to merge this into the neighbors function.

class GridWithWeights(SquareGrid):
    def __init__(self, width, height):
        super().__init__(width, height)
        self.weights = {}
    
    def cost(self, from_node, to_node):
        return self.weights.get(to_node, 1)

1.3.2 Queue with priorities

A priority queue associates with each item a number called a “priority”. When returning an item, it picks the one with the lowest number.

insert
Add item to queue
remove
Remove item with the lowest number
reprioritize
(optional) Change an existing item’s priority to a lower number

Here’s a reasonably fast priority queue that uses binary heaps, but does not support reprioritize. To get the right ordering, we’ll use tuples (priority, item). When an element is inserted that is already in the queue, we’ll have a duplicate; I’ll explain why that’s ok in the Optimization section.

import heapq

class PriorityQueue:
    def __init__(self):
        self.elements = []
    
    def empty(self):
        return len(self.elements) == 0
    
    def put(self, item, priority):
        heapq.heappush(self.elements, (priority, item))
    
    def get(self):
        return heapq.heappop(self.elements)[1]

1.3.3 Search

Here’s a tricky bit about the implementation: once we add movement costs it’s possible to visit a location again, with a better cost_so_far. That means the line if next not in came_from won’t work. Instead, have to check if the cost has gone down since the last time we visited. (In the original version of the article I wasn’t checking this, but my code worked anyway; I wrote some notes about that bug.)

This forest map is from the main page.

def dijkstra_search(graph, start, goal):
    frontier = PriorityQueue()
    frontier.put(start, 0)
    came_from = {}
    cost_so_far = {}
    came_from[start] = None
    cost_so_far[start] = 0
    
    while not frontier.empty():
        current = frontier.get()
        
        if current == goal:
            break
        
        for next in graph.neighbors(current):
            new_cost = cost_so_far[current] + graph.cost(current, next)
            if next not in cost_so_far or new_cost < cost_so_far[next]:
                cost_so_far[next] = new_cost
                priority = new_cost
                frontier.put(next, priority)
                came_from[next] = current
    
    return came_from, cost_so_far

Finally, after searching I need to build the path:

def reconstruct_path(came_from, start, goal):
    current = goal
    path = [current]
    while current != start:
        current = came_from[current]
        path.append(current)
    path.append(start) # optional
    path.reverse() # optional
    return path

Although paths are best thought of as a sequence of edges, it’s convenient to store them as a sequence of nodes. To build the path, start at the end and follow the came_from map, which points to the previous node. When we reach start, we’re done. It is the backwards path, so call reverse() at the end of reconstruct_path if you need it to be stored forwards. Sometimes it’s actually more convenient to store it backwards. Sometimes it’s useful to also store the start node in the list.

Let’s try it out:

from implementation import *
came_from, cost_so_far = dijkstra_search(diagram4, (1, 4), (7, 8))
draw_grid(diagram4, width=3, point_to=came_from, start=(1, 4), goal=(7, 8))
print()
draw_grid(diagram4, width=3, number=cost_so_far, start=(1, 4), goal=(7, 8))
print()
draw_grid(diagram4, width=3, path=reconstruct_path(came_from, start=(1, 4), goal=(7, 8)))

The line if next not in cost_so_far or new_cost < cost_so_far[next] could be simplified to if new_cost < cost_so_far.get(next, Infinity) but I didn’t want to explain Python’s get() in the main article so I left it as is. Another approach would be to use collections.defaultdict defaulting to infinity.

1.4 A* Search

Both Greedy Best-First Search and A* use a heuristic function. The only difference is that A* uses both the heuristic and the ordering from Dijkstra’s Algorithm. I’m going to show A* here.

def heuristic(a, b):
    (x1, y1) = a
    (x2, y2) = b
    return abs(x1 - x2) + abs(y1 - y2)

def a_star_search(graph, start, goal):
    frontier = PriorityQueue()
    frontier.put(start, 0)
    came_from = {}
    cost_so_far = {}
    came_from[start] = None
    cost_so_far[start] = 0
    
    while not frontier.empty():
        current = frontier.get()
        
        if current == goal:
            break
        
        for next in graph.neighbors(current):
            new_cost = cost_so_far[current] + graph.cost(current, next)
            if next not in cost_so_far or new_cost < cost_so_far[next]:
                cost_so_far[next] = new_cost
                priority = new_cost + heuristic(goal, next)
                frontier.put(next, priority)
                came_from[next] = current
    
    return came_from, cost_so_far

Let’s try it out:

from implementation import *
came_from, cost_so_far = a_star_search(diagram4, (1, 4), (7, 8))
draw_grid(diagram4, width=3, point_to=came_from, start=(1, 4), goal=(7, 8))
print()
draw_grid(diagram4, width=3, number=cost_so_far, start=(1, 4), goal=(7, 8))

1.4.1 Straighter paths

If you implement this code in your own project you might find that some of the paths aren’t as “straight” as you’d like. This is normal. When using grids, especially grids where every step has the same movement cost, you end up with ties: many paths have exactly the same cost. A* ends up picking one of the many short paths, and very often it doesn’t look good to you. The quick hack is to break the ties, but it’s not entirely satisfactory. The better approach is to change the map representation, which makes A* a lot faster, and also produces straighter, better looking paths. However, that only works for mostly-static maps where every step has the same movement cost. For the demos on my page, I’m using a quick hack, but it only works with my slow priority queue. If you switch to a faster priority queue you’ll need a different quick hack.

#C++ Implementation

Note: some of the sample code needs to include redblobgames/pathfinding/a-star/implementation.cpp to run. I am using C++11 for this code so some of it will need to be changed if you use an older version of the C++ standard.

2.1 Breadth First Search

First read the Python section; I have code here but I’m not including the same explanations. First, a generic graph class:

template<typename L>
struct SimpleGraph {
  typedef L Location;
  typedef typename vector<Location>::iterator iterator;
  unordered_map<Location, vector<Location> > edges;

  inline const vector<Location>  neighbors(Location id) {
    return edges[id];
  }
};

and the same example graph used in the Python code above, using char for location labels:

implementation-example-graph.png

SimpleGraph<char> example_graph {{
    {'A', {'B'}},
    {'B', {'A', 'C', 'D'}},
    {'C', {'A'}},
    {'D', {'E', 'A'}},
    {'E', {'B'}}
  }};

Instead of defining a queue class, I’ll use the one from the C++ standard library. Now we can try Breadth First Search:

#include "redblobgames/pathfinding/a-star/implementation.cpp"

template<typename Graph>
void breadth_first_search(Graph graph, typename Graph::Location start) {
  typedef typename Graph::Location Location;
  queue<Location> frontier;
  frontier.push(start);

  unordered_set<Location> visited;
  visited.insert(start);

  while (!frontier.empty()) {
    auto current = frontier.front();
    frontier.pop();

    std::cout << "Visiting " << current << std::endl;
    for (auto next : graph.neighbors(current)) {
      if (!visited.count(next)) {
        frontier.push(next);
        visited.insert(next);
      }
    }
  }
}


int main() {
  breadth_first_search(example_graph, 'A');
}

The code is slightly longer than the Python code, but it’s not too bad.

How about square grids? I’ll define another graph class. Note that it’s not related to the previous graph class. I’m using templates, not inheritance. The graph only needs to provide a Location typedef and a neighbors function.

struct SquareGrid {
  typedef tuple<int,int> Location;
  static array<Location, 4> DIRS;

  int width, height;
  unordered_set<Location> walls;

  SquareGrid(int width_, int height_)
     : width(width_), height(height_) {}

  inline bool in_bounds(Location id) const {
    int x, y;
    tie (x, y) = id;
    return 0 <= x && x < width && 0 <= y && y < height;
  }

  inline bool passable(Location id) const {
    return !walls.count(id);
  }

  vector<Location> neighbors(Location id) const {
    int x, y, dx, dy;
    tie (x, y) = id;
    vector<Location> results;

    for (auto dir : DIRS) {
      tie (dx, dy) = dir;
      Location next(x + dx, y + dy);
      if (in_bounds(next) && passable(next)) {
        results.push_back(next);
      }
    }

    if ((x + y) % 2 == 0) {
      // aesthetic improvement on square grids
      std::reverse(results.begin(), results.end());
    }

    return results;
  }
};

array<SquareGrid::Location, 4> SquareGrid::DIRS {Location{1, 0}, Location{0, -1}, Location{-1, 0}, Location{0, 1}};

In the helper file implementation.cpp I defined a function to draw grids:

#include "redblobgames/pathfinding/a-star/implementation.cpp"

int main() {
  SquareGrid grid = make_diagram1();
  draw_grid(grid, 2);
}

Let’s try Breadth First Search again, keeping track of came_from:

#include "redblobgames/pathfinding/a-star/implementation.cpp"

template<typename Graph>
unordered_map<typename Graph::Location, typename Graph::Location>
breadth_first_search(Graph graph, typename Graph::Location start) {
  typedef typename Graph::Location Location;
  queue<Location> frontier;
  frontier.push(start);

  unordered_map<Location, Location> came_from;
  came_from[start] = start;

  while (!frontier.empty()) {
    auto current = frontier.front();
    frontier.pop();

    for (auto next : graph.neighbors(current)) {
      if (!came_from.count(next)) {
        frontier.push(next);
        came_from[next] = current;
      }
    }
  }
  return came_from;
}

int main() {
  SquareGrid grid = make_diagram1();
  auto parents = breadth_first_search(grid, std::make_tuple(7, 8));
  draw_grid(grid, 2, nullptr, &parents);
}

Some implementations use internal storage, creating a Node object to hold came_from and other values for each graph node. I’ve instead chosen to use external storage, creating a single std::unordered_map to store the came_from for all graph nodes. If you know your map locations have integer indices, another option is to use a 1D or 2D array/vector to store came_from and other values.

2.1.1TODO Location struct

In this code I’m using C++ std::tuple because I used tuples in my Python code. However, tuples are unfamiliar to most C++ programmers. It’d be a little more verbose to define a struct with {x,y}, because I would need to define the constructor, copy constructor, assignment operator, and equality comparison, along with a hash function, but that kind of code is familar to most C++ programmers. I should change it.

Another option (which I use in my own code) is to encode {x,y} in an int. If the A* code always has a dense set of integers instead of arbitrary Location types, it enables some further optimizations. We can use an array for the various sets instead of hash tables. We can leave most of them uninitialized. For the arrays that have to be reinitialized each time, we can make them persistent across calls to A* (maybe using thread-local storage), reinitializing only the parts of the array that were used on the previous call. These are all more advanced techniques that I don’t want to use in a beginner-level tutorial.

2.2 Early Exit

As with the Python version, all we have to do is add a parameter to the function and a test to the main loop:

#include "redblobgames/pathfinding/a-star/implementation.cpp"

template<typename Graph>
unordered_map<typename Graph::Location, typename Graph::Location>
breadth_first_search(const Graph& graph,
                     typename Graph::Location start,
                     typename Graph::Location goal) {
  typedef typename Graph::Location Location;
  queue<Location> frontier;
  frontier.push(start);

  unordered_map<Location, Location> came_from;
  came_from[start] = start;

  while (!frontier.empty()) {
    auto current = frontier.front();
    frontier.pop();

    if (current == goal) {
      break;
    }
    
    for (auto next : graph.neighbors(current)) {
      if (!came_from.count(next)) {
        frontier.push(next);
        came_from[next] = current;
      }
    }
  }
  return came_from;
}

int main() {
  SquareGrid grid = make_diagram1();
  auto parents = breadth_first_search(grid, SquareGrid::Location{8, 7}, SquareGrid::Location{17, 2});
  draw_grid(grid, 2, nullptr, &parents);
}

2.3 Dijkstra’s Algorithm

2.3.1 Graph with weights

Here’s a grid with a list of forest tiles, which will have movement cost 5. In this forest map I chose to make movement depend only on to_node, but there are other types of movement that use both nodes.

struct GridWithWeights: SquareGrid {
  unordered_set<Location> forests;
  GridWithWeights(int w, int h): SquareGrid(w, h) {}
  double cost(Location from_node, Location to_node) const {
    return forests.count(to_node) ? 5 : 1;
  }
};

2.3.2 Queue with priorities

We need a priority queue. C++ offers a priority_queue class that uses a binary heap but not the reprioritize operation. I’ll use a pair (priority, item) for the queue elements to get the right ordering. By default, the C++ priority queue returns the maximum element first, using the std::less comparator; we want the minimum element instead, so I’ll use the std::greater comparator.

template<typename T, typename priority_t>
struct PriorityQueue {
  typedef pair<priority_t, T> PQElement;
  priority_queue<PQElement, vector<PQElement>,
                 std::greater<PQElement>> elements;

  inline bool empty() const { return elements.empty(); }

  inline void put(T item, priority_t priority) {
    elements.emplace(priority, item);
  }

  inline T get() {
    T best_item = elements.top().second;
    elements.pop();
    return best_item;
  }
};

In this sample code I’m wrapping the C++ std::priority_queue class but I think it’d be reasonable to use that class directly without the wrapper.

2.3.3 Search

See the forest map from the main page.

template<typename Graph>
void dijkstra_search
  (const Graph& graph,
   typename Graph::Location start,
   typename Graph::Location goal,
   unordered_map<typename Graph::Location, typename Graph::Location>& came_from,
   unordered_map<typename Graph::Location, double>& cost_so_far)
{
  typedef typename Graph::Location Location;
  PriorityQueue<Location, double> frontier;
  frontier.put(start, 0);

  came_from[start] = start;
  cost_so_far[start] = 0;
  
  while (!frontier.empty()) {
    auto current = frontier.get();

    if (current == goal) {
      break;
    }

    for (auto next : graph.neighbors(current)) {
      double new_cost = cost_so_far[current] + graph.cost(current, next);
      if (!cost_so_far.count(next) || new_cost < cost_so_far[next]) {
        cost_so_far[next] = new_cost;
        came_from[next] = current;
        frontier.put(next, new_cost);
      }
    }
  }
}

The types of the cost variables should all match the types used in the graph. If you use int then you can use int for the cost variable and the priorities in the priority queue; if you use double then you should use double for these. In this code I used double but I could’ve used int and it would’ve worked the same. However, if your graph edge costs are doubles or if your heuristic uses doubles, then you’ll need to use doubles here.

Finally, after searching I need to build the path:

template<typename Location>
vector<Location> reconstruct_path(
   Location start,
   Location goal,
   unordered_map<Location, Location>& came_from
) {
  vector<Location> path;
  Location current = goal;
  path.push_back(current);
  while (current != start) {
    current = came_from[current];
    path.push_back(current);
  }
  path.push_back(start); // optional
  std::reverse(path.begin(), path.end());
  return path;
}

Although paths are best thought of as a sequence of edges, it’s convenient to store them as a sequence of nodes. To build the path, start at the end and follow the came_from map, which points to the previous node. When we reach start, we’re done. It is the backwards path, so call reverse() at the end of reconstruct_path if you need it to be stored forwards. Sometimes it’s actually more convenient to store it backwards. Sometimes it’s useful to also store the start node in the list.

Let’s try it out:

#include "redblobgames/pathfinding/a-star/implementation.cpp"

int main() {
  GridWithWeights grid = make_diagram4();
  SquareGrid::Location start{1, 4};
  SquareGrid::Location goal{8, 5};
  unordered_map<SquareGrid::Location, SquareGrid::Location> came_from;
  unordered_map<SquareGrid::Location, double> cost_so_far;
  dijkstra_search(grid, start, goal, came_from, cost_so_far);
  draw_grid(grid, 2, nullptr, &came_from);
  std::cout << std::endl;
  draw_grid(grid, 3, &cost_so_far, nullptr);
  std::cout << std::endl;
  vector<SquareGrid::Location> path = reconstruct_path(start, goal, came_from);
  draw_grid(grid, 3, nullptr, nullptr, &path);
}

The results are not exactly the same as the Python version because I’m using the built-in hash tables for the priority queue, and the order of iteration over hash tables is not going to be consistent.

2.4 A* Search

A* is almost exactly like Dijkstra’s Algorithm, except we add in a heuristic. Note that the code for the algorithm isn’t specific to grids. Knowledge about grids is in the graph class (SquareGrids in this case) and in the heuristic function. Replace those two and you can use the A* algorithm code with any other graph structure.

inline double heuristic(SquareGrid::Location a, SquareGrid::Location b) {
  int x1, y1, x2, y2;
  tie (x1, y1) = a;
  tie (x2, y2) = b;
  return abs(x1 - x2) + abs(y1 - y2);
}

template<typename Graph>
void a_star_search
  (const Graph& graph,
   typename Graph::Location start,
   typename Graph::Location goal,
   unordered_map<typename Graph::Location, typename Graph::Location>& came_from,
   unordered_map<typename Graph::Location, double>& cost_so_far)
{
  typedef typename Graph::Location Location;
  PriorityQueue<Location, double> frontier;
  frontier.put(start, 0);

  came_from[start] = start;
  cost_so_far[start] = 0;
  
  while (!frontier.empty()) {
    auto current = frontier.get();

    if (current == goal) {
      break;
    }

    for (auto next : graph.neighbors(current)) {
      double new_cost = cost_so_far[current] + graph.cost(current, next);
      if (!cost_so_far.count(next) || new_cost < cost_so_far[next]) {
        cost_so_far[next] = new_cost;
        double priority = new_cost + heuristic(next, goal);
        frontier.put(next, priority);
        came_from[next] = current;
      }
    }
  }
}

The type of the priority values including the type used in the priority queue should be big enough to include both the graph costs (cost_t) and the heuristic value. For example, if the graph costs are ints and the heuristic returns a double, then you need the priority queue to accept doubles. In this sample code I use double for all three (cost, heuristic, and priority), but I could’ve used int because my costs and heuristics are integer valued.

Minor note: It would be more correct to write frontier.put(start, heuristic(start, goal)) than frontier.put(start, 0) but it makes no difference here because the start node’s priority doesn’t matter. It is the only node in the priority queue and it is selected and removed before anything else is put in there.

#include "redblobgames/pathfinding/a-star/implementation.cpp"

int main() {
  GridWithWeights grid = make_diagram4();
  SquareGrid::Location start{1, 4};
  SquareGrid::Location goal{8, 5};
  unordered_map<SquareGrid::Location, SquareGrid::Location> came_from;
  unordered_map<SquareGrid::Location, double> cost_so_far;
  a_star_search(grid, start, goal, came_from, cost_so_far);
  draw_grid(grid, 2, nullptr, &came_from);
  std::cout << std::endl;
  draw_grid(grid, 3, &cost_so_far, nullptr);
  std::cout << std::endl;
  vector<SquareGrid::Location> path = reconstruct_path(start, goal, came_from);
  draw_grid(grid, 3, nullptr, nullptr, &path);
}

2.4.1 Straighter paths

If you implement this code in your own project you might find that some of the paths aren’t as “straight” as you’d like. This is normal. When using grids, especially grids where every step has the same movement cost, you end up with ties: many paths have exactly the same cost. A* ends up picking one of the many short paths, and very often it doesn’t look good to you. The quick hack is to break the ties, but it’s not entirely satisfactory. The better approach is to change the map representation, which makes A* a lot faster, and also produces straighter, better looking paths. However, that only works for mostly-static maps where every step has the same movement cost. For the demos on my page, I’m using a quick hack, but it only works with my slow priority queue. If you switch to a faster priority queue you’ll need a different quick hack.

2.4.2TODO Heuristic function should be template parameter

I should make the heuristic into a template parameter instead of being a global function.

#C# Implementation

These were my first C# programs so they might not be idiomatic or stylistically proper. These examples aren’t as complete as the Python and C++ sections, but I hope they’re helpful.

Here’s a simple graph, and Breadth First Search:

using System;
using System.Collections.Generic;


public class Graph<Location>
{
    // NameValueCollection would be a reasonable alternative here, if
    // you're always using string location types
    public Dictionary<Location, Location[]> edges
        = new Dictionary<Location, Location[]>();

    public Location[] Neighbors(Location id)
    {
        return edges[id];
    }
};


class BreadthFirstSearch
{
    static void Search(Graph<string> graph, string start)
    {
        var frontier = new Queue<string>();
        frontier.Enqueue(start);

        var visited = new HashSet<string>();
        visited.Add(start);

        while (frontier.Count > 0)
        {
            var current = frontier.Dequeue();

            Console.WriteLine("Visiting {0}", current);
            foreach (var next in graph.Neighbors(current))
            {
                if (!visited.Contains(next)) {
                    frontier.Enqueue(next);
                    visited.Add(next);
                }
            }
        }
    }
    
    static void Main()
    {
        Graph<string> g = new Graph<string>();
        g.edges = new Dictionary<string, string[]>
            {
            { "A", new [] { "B" } },
            { "B", new [] { "A", "C", "D" } },
            { "C", new [] { "A" } },
            { "D", new [] { "E", "A" } },
            { "E", new [] { "B" } }
        };

        Search(g, "A");
    }
}

Here’s a graph representing a grid with weighted edges (the forest and walls example from the main page):

using System;
using System.Collections.Generic;


// A* needs only a WeightedGraph and a location type L, and does *not*
// have to be a grid. However, in the example code I am using a grid.
public interface WeightedGraph<L>
{
    double Cost(Location a, Location b);
    IEnumerable<Location> Neighbors(Location id);
}


public struct Location
{
    // Implementation notes: I am using the default Equals but it can
    // be slow. You'll probably want to override both Equals and
    // GetHashCode in a real project.
    
    public readonly int x, y;
    public Location(int x, int y)
    {
        this.x = x;
        this.y = y;
    }
}


public class SquareGrid : WeightedGraph<Location>
{
    // Implementation notes: I made the fields public for convenience,
    // but in a real project you'll probably want to follow standard
    // style and make them private.
    
    public static readonly Location[] DIRS = new []
        {
            new Location(1, 0),
            new Location(0, -1),
            new Location(-1, 0),
            new Location(0, 1)
        };

    public int width, height;
    public HashSet<Location> walls = new HashSet<Location>();
    public HashSet<Location> forests = new HashSet<Location>();

    public SquareGrid(int width, int height)
    {
        this.width = width;
        this.height = height;
    }

    public bool InBounds(Location id)
    {
        return 0 <= id.x && id.x < width
            && 0 <= id.y && id.y < height;
    }

    public bool Passable(Location id)
    {
        return !walls.Contains(id);
    }

    public double Cost(Location a, Location b)
    {
        return forests.Contains(b) ? 5 : 1;
    }
    
    public IEnumerable<Location> Neighbors(Location id)
    {
        foreach (var dir in DIRS) {
            Location next = new Location(id.x + dir.x, id.y + dir.y);
            if (InBounds(next) && Passable(next)) {
                yield return next;
            }
        }
    }
}


public class PriorityQueue<T>
{
    // I'm using an unsorted array for this example, but ideally this
    // would be a binary heap. Find a binary heap class:
    // * https://bitbucket.org/BlueRaja/high-speed-priority-queue-for-c/wiki/Home
    // * http://visualstudiomagazine.com/articles/2012/11/01/priority-queues-with-c.aspx
    // * http://xfleury.github.io/graphsearch.html
    // * http://stackoverflow.com/questions/102398/priority-queue-in-net
    
    private List<Tuple<T, double>> elements = new List<Tuple<T, double>>();

    public int Count
    {
        get { return elements.Count; }
    }
    
    public void Enqueue(T item, double priority)
    {
        elements.Add(Tuple.Create(item, priority));
    }

    public T Dequeue()
    {
        int bestIndex = 0;

        for (int i = 0; i < elements.Count; i++) {
            if (elements[i].Item2 < elements[bestIndex].Item2) {
                bestIndex = i;
            }
        }

        T bestItem = elements[bestIndex].Item1;
        elements.RemoveAt(bestIndex);
        return bestItem;
    }
}


/* NOTE about types: in the main article, in the Python code I just
 * use numbers for costs, heuristics, and priorities. In the C++ code
 * I use a typedef for this, because you might want int or double or
 * another type. In this C# code I use double for costs, heuristics,
 * and priorities. You can use an int if you know your values are
 * always integers, and you can use a smaller size number if you know
 * the values are always small. */

public class AStarSearch
{
    public Dictionary<Location, Location> cameFrom
        = new Dictionary<Location, Location>();
    public Dictionary<Location, double> costSoFar
        = new Dictionary<Location, double>();

    // Note: a generic version of A* would abstract over Location and
    // also Heuristic
    static public double Heuristic(Location a, Location b)
    {
        return Math.Abs(a.x - b.x) + Math.Abs(a.y - b.y);
    }

    public AStarSearch(WeightedGraph<Location> graph, Location start, Location goal)
    {
        var frontier = new PriorityQueue<Location>();
        frontier.Enqueue(start, 0);

        cameFrom[start] = start;
        costSoFar[start] = 0;

        while (frontier.Count > 0)
        {
            var current = frontier.Dequeue();

            if (current.Equals(goal))
            {
                break;
            }

            foreach (var next in graph.Neighbors(current))
            {
                double newCost = costSoFar[current]
                    + graph.Cost(current, next);
                if (!costSoFar.ContainsKey(next)
                    || newCost < costSoFar[next])
                {
                    costSoFar[next] = newCost;
                    double priority = newCost + Heuristic(next, goal);
                    frontier.Enqueue(next, priority);
                    cameFrom[next] = current;
                }
            }
        }
    }
}

public class Test
{
    static void DrawGrid(SquareGrid grid, AStarSearch astar) {
        // Print out the cameFrom array
        for (var y = 0; y < 10; y++)
        {
            for (var x = 0; x < 10; x++)
            {
                Location id = new Location(x, y);
                Location ptr = id;
                if (!astar.cameFrom.TryGetValue(id, out ptr))
                {
                    ptr = id;
                }
                if (grid.walls.Contains(id)) { Console.Write("##"); }
                else if (ptr.x == x+1) { Console.Write("\u2192 "); }
                else if (ptr.x == x-1) { Console.Write("\u2190 "); }
                else if (ptr.y == y+1) { Console.Write("\u2193 "); }
                else if (ptr.y == y-1) { Console.Write("\u2191 "); }
                else { Console.Write("* "); }
            }
            Console.WriteLine();
        }
    }
    
    static void Main()
    {
        // Make "diagram 4" from main article
        var grid = new SquareGrid(10, 10);
        for (var x = 1; x < 4; x++)
        {
            for (var y = 7; y < 9; y++)
            {
                grid.walls.Add(new Location(x, y));
            }
        }
        grid.forests = new HashSet<Location>
            {
                new Location(3, 4), new Location(3, 5),
                new Location(4, 1), new Location(4, 2),
                new Location(4, 3), new Location(4, 4),
                new Location(4, 5), new Location(4, 6),
                new Location(4, 7), new Location(4, 8),
                new Location(5, 1), new Location(5, 2),
                new Location(5, 3), new Location(5, 4),
                new Location(5, 5), new Location(5, 6),
                new Location(5, 7), new Location(5, 8),
                new Location(6, 2), new Location(6, 3),
                new Location(6, 4), new Location(6, 5),
                new Location(6, 6), new Location(6, 7),
                new Location(7, 3), new Location(7, 4),
                new Location(7, 5)
            };

        // Run A*
        var astar = new AStarSearch(grid, new Location(1, 4),
                                    new Location(8, 5));

        DrawGrid(grid, astar);
    }
}

#Optimizations

For the code I present here, I’ve been focusing on simplicity and generality rather than performance. First make it work, then make it fast. Many of the optimizations I use in real projects are specific to the project, so instead of presenting optimal code, here are some ideas to pursue for your own project:

4.1 Graph

The biggest optimization you can make is to explore fewer nodes. My #1 recommendation is that if you’re using a grid map, consider using a non-grid pathfinding graph. It’s not always feasible but it’s worth looking at.

If your graph has a simple structure (e.g. a grid), calculate the neighbors in a function. If it’s a more complex structure (either a non-grid, or a grid with lots of walls, like a maze), store the neighbors in a data structure.

4.2 Queue

Breadth First Search uses a simple queue instead of the priority queue needed by the other algorithms. Queues are simpler and faster than priority queues. In exchange, the other algorithms usually explore fewer nodes. In most game maps, exploring fewer nodes is worth the slowdown from the other algorithms. There are some maps though where you don’t save much, and it might be better to use Breadth First Search.

For queues, use a deque instead of an array. A deque allows fast insertion and removal on either end, whereas an array is fast only at one end. In Python, see collections.deque; in C++, see the deque container.

For priority queues, use a binary heap instead of an array or sorted array. A binary heap allows fast insertion and removal, whereas an array is fast at one or the other but not both. In Python, see heapq; in C++, see the priority_queue container.

In Python, the Queue and PriorityQueue classes I presented above are so simple that you might consider inlining the methods into the search algorithm. I don’t know if this buys you much; I need to measure it. The C++ versions are going to be inlined.

In Dijkstra’s Algorithm, note that the priority queue’s priority is stored twice, once in the priority queue and once in cost_so_far, so you could write a priority queue that gets priorities from elsewhere. I’m not sure if it’s worth it.

The standard way Dijkstra’s Algorithm is presented uses a reprioritizing priority queue. However, what happens if you don’t reprioritize? You’ll end up with duplicate elements in there. The algorithm still works. You’ll revisit some locations more than you need to. The priority queue will have more elements than necessary, which slows it down, but the data structures that support reprioritization also slow down by having more elements. See the paper “Priority Queues and Dijkstra’s Algorithm” by Chen, Chowdhury, Ramachandran, Lan Roche, Tong. That paper also suggests looking at pairing heaps and other data structures.

If you’re considering using something other than a binary heap, first measure the size of your frontier and how often you reprioritize. Profile the code and see if the priority queue is the bottleneck.

My gut feeling is that bucketing is promising. Just as bucket sort and radix sort can be useful alternatives to quicksort when the keys are integers, we have an even better situation with Dijkstra’s Algorithm and A*. The priorities in Dijkstra’s Algorithm are incredibly narrow. If the lowest element in the queue has priority f, then the highest element has priority f+e where e is the maximum edge weight. In the forest example, I have edge weights 1 and 5. That means all the priorities in the queue are going to be between f and f+5. Since they’re all integers, there are only six different priorities. We could use six buckets and not sort anything at all! A* produces a wider range of priorities but it’s still worth looking at. And there are fancier bucketing approaches that handle a wider range of situations.

I have more note about priority queue data structures here.

4.3 Search

The heuristic adds complexity and cpu time. The goal though is to explore fewer nodes. In some maps (such as mazes), the heuristic may not add much information, and it may be better to use a simpler algorithm without a heuristic guide.

If your graph uses integers as locations, consider using a simple array instead of a hash table for cost_so_far, visited, came_from, etc. Since visited is an array of booleans, you can use a bit vector. Initialize the bit vector for all ids, but leave cost_so_far and came_from uninitialized (if your language allows). Then only initialize on the first visit.

If you run only one search at a time, you can statically allocate and then reuse data structures from one invocation to the next. Instead of initializing them on entry, reset them on exit. You can use an array to keep track of which locations were modified, then reset only those. For example, if you’re using a visited[] array initialized for 1000 nodes, but most searches visit fewer than 100 nodes, then you can keep an array of indices modified, and then reset only those 100 nodes on exiting the function instead of re-initializing all 1000 nodes when you enter the function.

Some people use an inadmissible (overestimating) heuristic to speed up A* search. This seems reasonable. I haven’t looked closely into its implications though. I believe (but don’t know for sure) that some already-visited elements may need to be visited again even after they’ve been taken out of the frontier.

Some implementations always insert a new node into the open set, even if it’s already there. You can avoid the potentially expensive step of checking whether the node is already in the open set. This will make your open set bigger/slower and you’ll also end up evaluating more nodes than necessary. If the open-set test is expensive, it might still be worth it. However, in the code I’ve presented, I made the test cheap and I don’t use this approach.

Some implementations don’t test whether a new node is better than an existing node in the open set. This avoids a potentially expensive check. However, it also can lead to a bug. For some types of maps, you will not find the shortest path when you skip this test. In the code I’ve presented, I check this (new_cost < cost_so_far). The test is cheap because I made it cheap to look up cost_so_far.

#Troubleshooting

5.1 Wrong path

If you’re not getting a shortest path, try testing:

  • Does your priority queue work correctly? Try stopping the search and dequeuing all the elements. They should all be in order.
  • Does your heuristic ever overestimate the true distance? The priority of a new node should never be lower than the priority of its parent, unless you are overestimating the distance (you can do this but you won’t get shortest paths anymore).
  • In a statically typed language, the cost, heuristic, and priority values need to have compatible types. The sample code on this page works with either integers or floating point types, but not all graphs and heuristics are limited to integer values. Since priorities are the sum of costs and heuristics, the priorities will need to be floating point if either costs or heuristics are floating point.

5.2 Ugly path

The most common question I get when people run A* on a grid is why don’t my paths look straight? If you’ve told A* that all grid movements are equal, then there are lots of shortest paths of the same length, and it’s going to pick one arbitrarily. The path is short but it doesn’t look good.

  • One solution is to straighten the paths afterwards, using a “string pulling” algorithm.
  • One solution is to guide the paths in the right direction, by adjusting the heuristic. There are some cheap tricks that don’t work in all situations; read more here.
  • One solution is to not use a grid. Tell A* just the places where you might turn, instead of every grid square; read more here.
  • One solution is a hack, but it works some of the time. When iterating through neighbors, instead of always using the same ordering (such as north, east, south, west), change the ordering on “odd” grid nodes (those where (x+y) % 2 == 1). I use this trick on these tutorial pages.

#More reading

  • Algorithms textbooks often use mathematical notation with single-letter variable names. In these pages I’ve tried to use more descriptive variable names. Correspondences:
    • cost is sometimes written as w or d or l or length
    • cost_so_far is usually written as g or d or distance
    • heuristic is usually written as h
    • In A*, the priority is usually written as f, where f = g + h
    • came_from is sometimes written as π or parent or previous or prev
    • frontier is usually called OPEN
    • visited is the union of OPEN and CLOSED
    • locations such as current and next are written with letters u, v
  • Wikipedia links: