blog/algorithms/11-paths/2024-01-01-bf-to-astar/01-bf.md
Matej Focko 2cf4a3efba
chore: run pre-commit
Signed-off-by: Matej Focko <me@mfocko.xyz>
2024-01-03 19:38:35 +01:00

18 KiB
Raw Permalink Blame History

id slug title description tags last_update
bf /paths/bf-to-astar/bf BF Solving the shortest path problem with a naïve approach that turns into something.
cpp
brute force
bellman ford
dynamic programming
date
2024-01-01

Basic idea

We will ease in with our own algorithm to find the shortest path. We will start by thinking about the ways we can achieve that. If we didn't have the * cells, we could've easily run a BFS1 and be done with it. Maybe it is a good place to start, or isn't, there is only one way to find out though.

How does the BFS work? We know the vertex where we start and we know the vertex we want to find the shortest path to. Given this knowledge we incrementally visit all of our neighbours and we do that over and over until the destination is found2. Could we leverage this somehow?

Naïve approach

Well, we could probably start with all vertices being unreachable (having the highest possible price) and try to improve what we've gotten so far until there are no improvements. That sounds fine, we shall implement this. Since we are going on repeat, we will name this function bf() as in brute-force, cause it is trying to find it the hard way:

const static std::vector<vertex_t> DIRECTIONS =
    std::vector{std::make_pair(0, 1), std::make_pair(0, -1),
                std::make_pair(1, 0), std::make_pair(-1, 0)};

auto bf(const graph& g, const vertex_t& source, const vertex_t& destination)
    -> int {
  // source must be within the bounds
  assert(g.has(source));

  // destination must be within the bounds
  assert(g.has(destination));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we need to improve the paths as long as possible
  bool improvement_found;
  do {
    // reset the flag at the beginning
    improvement_found = false;

    // go through all of the vertices
    for (int y = 0; y < g.height(); ++y) {
      for (int x = 0; x < g.width(); ++x) {
        // skip the cells we cannot reach
        if (distances[y][x] == graph::unreachable()) {
          continue;
        }

        // go through the neighbours
        auto u = std::make_pair(x, y);
        for (const auto& [dx, dy] : DIRECTIONS) {
          auto v = std::make_pair(x + dx, y + dy);
          auto cost = g.cost(u, v);

          // if we can move to the cell and it's better, relax¹ it
          if (cost != graph::unreachable() &&
              distances[y][x] + cost < distances[y + dy][x + dx]) {
            distances[y + dy][x + dx] = distances[y][x] + cost;
            improvement_found = true;
          }
        }
      }
    }
  } while (improvement_found);

  return distances[destination.second][destination.first];
}

:::info Relaxation

I have made a brief mention of the relaxation in the comment in the code. You've been probably thought that relaxation of an edge means that you found a better solution to the problem.

In general it is an approximation technique that reduces the problem of finding the path u → x1 → … → xn → v to subproblems u → x1, x1 → x2, …, xn → v such that the sum of the costs of each step is minimal.

:::

Correctness

Is our solution correct? It appears to be correct… We have rather complicated map and our algorithm has finished in an instant with the following output:

Normal cost: 1
Vortex cost: 5
Graph:
#############
#..#..*.*.**#
##***.....**#
#..########.#
#...###...#.#
#..#...##.#.#
#..#.*.#..#.#
#D...#....#.#
########*.*.#
#S..........#
#############
Cost: 22

If you have a better look at the map, you will realize that the cost 22 is the one path skipping the * cells, since they cost more than going around.

We can play around a bit with it. The * cells can be even vortices that pull you in with a negative price and let you propel yourself out 😉 Let's change their cost to -1 then. Let's check what's the fastest path to the cell.

Normal cost: 1
Vortex cost: -1
Graph:
#############
#..#..*.*.**#
##***.....**#
#..########.#
#...###...#.#
#..#...##.#.#
#..#.*.#..#.#
#D...#....#.#
########*.*.#
#S..........#
#############

And we're somehow stuck… The issue comes from the fact that spinning around in the vortices allows us to lower the cost infinitely. That's why after each iteration there is still a possibility to lower the cost, hence the algorithm doesn't finish. What can we do about this?

:::tip

This algorithm is correct as long as there are no negative loops, i.e. ways how to lower the cost infinitely. Therefore we can also just lay a precondition that requires no negative loops to be present.

:::

Fixing the infinite loop

Our issue lies in the fact that we can endlessly lower the cost. Such thing must surely happen in some kind of a loop. We could probably track the relaxations and once we spot repeating patterns, we know we can safely terminate with some results at least.

This approach will not even work on our 2D map, let alone any graph. Problem is that the negative loops lower the cost in each iteration and that results in lowering of the costs to the cells that are reachable from the said loops. That's why this problem is relatively hard to tackle, it's not that easy to spot the repeating patterns algorithmically.

On the other hand, we can approach this from the different perspective. Let's assume the worst-case scenario (generalized for any graph):

Let K_n be complete graph. Let P be the shortest path from v_1 to v_n such that P has n - 1 edges, i.e. the shortest path between the two chosen vertices visits all vertices (not necessarily in order) and has the lowest cost.

In such scenario assume the worst-case ordering of the relaxations (only one helpful relaxation per iteration). In this case, in each iteration we find the next edge on our path P as the last. This means that we need \vert V \vert - 1 iterations to find the shortest path P.

Because we have laid P as the shortest path from v_1 to v_n and it visits all vertices, its prefixes are the shortest paths from v_1 to any other vertex in our graph.

Therefore, we can safely assume that any relaxation after \vert V \vert - 1 iterations, is the effect of a negative loop in the graph.

How can we leverage this? We will go through the edges only as many times as cells we have. Let's adjust the code to fix the looping:

auto bf_finite(const graph& g, const vertex_t& source,
               const vertex_t& destination) -> int {
  // source must be within the bounds
  assert(g.has(source));

  // destination must be within the bounds
  assert(g.has(destination));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we only iterate as many times as cells that we have
  for (int i = g.height() * g.width(); i > 0; --i) {
    // go through all of the vertices
    for (int y = 0; y < g.height(); ++y) {
      for (int x = 0; x < g.width(); ++x) {
        // skip the cells we cannot reach
        if (distances[y][x] == graph::unreachable()) {
          continue;
        }

        // go through the neighbours
        auto u = std::make_pair(x, y);
        for (const auto& [dx, dy] : DIRECTIONS) {
          auto v = std::make_pair(x + dx, y + dy);
          auto cost = g.cost(u, v);

          // if we can move to the cell and it's better, relax¹ it
          if (cost != graph::unreachable() &&
              distances[y][x] + cost < distances[y + dy][x + dx]) {
            distances[y + dy][x + dx] = distances[y][x] + cost;
          }
        }
      }
    }
  }

  return distances[destination.second][destination.first];
}

And we get the following result:

Normal cost: 1
Vortex cost: -1
Graph:
#############
#..#..*.*.**#
##***.....**#
#..########.#
#...###...#.#
#..#...##.#.#
#..#.*.#..#.#
#D...#....#.#
########*.*.#
#S..........#
#############
Cost: -236

The negative cost means that there is a way to propel ourselves via some vortices. Let's adjust the cost of vortices back to the original 5 and check whether our modified algorithm works as it did before. And it surely does yield the 22 as before.

:::tip Refactoring

You can definitely notice some deep nesting in our code, to counter this phenomenon I will convert the looping over x and y to one variable that can be decomposed to x and y. It is a very common practice when working with 2D arrays/lists to represent them as 1D. In our case:

i : 0 → width * height - 1
x = i % width
y = i / width

:::

Bellman-Ford

If you have ever attended any Algorithms course that had path-finding in its syllabus, you probably feel like you've seen the algorithm above before3… And yes, the first algorithm I have proposed is a very dumb version of the Bellman-Ford algorithm, it's dumb, because it loops 😉 After our “looping” prevention we got to the point that is almost the Bellman-Ford with the one exception that it doesn't report whether there are any negative cycles, it just ends.

Let's have a look at a proper implementation of the Bellman-Ford algorithm:

auto bellman_ford(const graph& g, const vertex_t& source)
    -> std::vector<std::vector<int>> {
  // source must be within the bounds
  assert(g.has(source));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we only iterate as many times as cells that we have
  for (int i = g.height() * g.width(); i > 0; --i) {
    // go through all of the vertices
    for (int v = g.height() * g.width() - 1; v >= 0; --v) {
      int y = v / g.width();
      int x = v % g.width();

      // skip the cells we cannot reach
      if (distances[y][x] == graph::unreachable()) {
        continue;
      }

      // go through the neighbours
      auto u = std::make_pair(x, y);
      for (const auto& [dx, dy] : DIRECTIONS) {
        auto v = std::make_pair(x + dx, y + dy);
        auto cost = g.cost(u, v);

        // if we can move to the cell and it's better, relax¹ it
        if (cost != graph::unreachable() &&
            distances[y][x] + cost < distances[y + dy][x + dx]) {
          distances[y + dy][x + dx] = distances[y][x] + cost;
        }
      }
    }
  }

  // now we check for the negative loops
  bool relaxed = false;
  for (int v = g.height() * g.width() - 1; !relaxed && v >= 0; --v) {
    int y = v / g.width();
    int x = v % g.width();

    // skip the cells we cannot reach
    if (distances[y][x] == graph::unreachable()) {
      continue;
    }

    // go through the neighbours
    auto u = std::make_pair(x, y);
    for (const auto& [dx, dy] : DIRECTIONS) {
      auto v = std::make_pair(x + dx, y + dy);
      auto cost = g.cost(u, v);

      // if we can move to the cell and it's better, relax¹ it
      if (cost != graph::unreachable() &&
          distances[y][x] + cost < distances[y + dy][x + dx]) {
        relaxed = true;
        std::cerr << "Found a negative loop\n";
        break;
      }
    }
  }

  return distances;
}

And if we run it with our negative cost of entering vortices:

[Bellman-Ford] Found a negative loop
[Bellman-Ford] Cost: -240

On the Bellman-Ford

You might be surprised that we have managed to iterate from a brute-force method that mindlessly tries to find a better path until there are no better paths left all the way to the Bellman-Ford algorithm.

I always say that Bellman-Ford is a smart brute-force. BF is also an algorithm that leverages dynamic programming. You might wonder how can it utilize DP if it is “technically” a brute-force technique. Table with the shortest distances is the thing that makes it DP.

I might not know the shortest path yet, but I do remember all of other paths, and I can improve them, if possible.

That's where the beauty of both dynamic programming and relaxing gets merged together and does its magic.

Proof of the correctness of the BF is done via induction to the number of iterations. I would suggest to try to prove the correctness yourself and possibly look it up, if necessary.

Also the correctness of the BF relies on the conclusion we've made when fixing the infinite-loop on our naïve BF solution.

Time complexity

Let's have a short look at the time complexities of the presented algorithms:

  1. naïve approach: given that there are no negative loops, we are bound by the worst-case ordering of the relaxations which results in

    \mathcal{O}(\vert V \vert \cdot \vert E \vert)

  2. our naïve approach with the fixed count of iterations instead of the do-while loop results in the same worst-case time complexity:

    \Theta(\vert V \vert \cdot \vert E \vert)

  3. and finally the well-known Bellman-Ford's algorithm time complexity: \Theta(\vert V \vert \cdot \vert E \vert)

Small refactor

Since we are literally copy-pasting the body of the loops just for the sake of relaxing, we can factor that part out into a separate function:

static auto _check_vertex(const graph& g,
                          std::vector<std::vector<int>>& distances, int v,
                          bool check_only = false) -> bool {
  bool improvement_found = false;

  // unpack the vertex coordinates
  int y = v / g.width();
  int x = v % g.width();

  // skip the cells we cannot reach
  if (distances[y][x] == graph::unreachable()) {
    return false;
  }

  // go through the neighbours
  auto u = std::make_pair(x, y);
  for (const auto& [dx, dy] : DIRECTIONS) {
    auto v = std::make_pair(x + dx, y + dy);
    auto cost = g.cost(u, v);

    // if we can move to the cell and it's better, relax¹ it
    if (cost != graph::unreachable() &&
        distances[y][x] + cost < distances[y + dy][x + dx]) {
      if (check_only) {
        return true;
      }

      distances[y + dy][x + dx] = distances[y][x] + cost;
      improvement_found = true;
    }
  }

  return improvement_found;
}

This function can be also used for checking the negative loops at the end of the BF by using the check_only parameter to signal that we just want to know if there would be any edge relaxed instead of performing the relaxation itself.

Then we can also see the differences between the specific versions of our path-finding algorithms in a clear way:

auto bf(const graph& g, const vertex_t& source, const vertex_t& destination)
    -> int {
  // source must be within the bounds
  assert(g.has(source));

  // destination must be within the bounds
  assert(g.has(destination));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we need to improve the paths as long as possible
  bool improvement_found;
  do {
    // reset the flag at the beginning
    improvement_found = false;

    // go through all of the vertices
    for (int v = g.height() * g.width() - 1; v >= 0; --v) {
      improvement_found = _check_vertex(g, distances, v) || improvement_found;
    }
  } while (improvement_found);

  return distances[destination.second][destination.first];
}

auto bf_finite(const graph& g, const vertex_t& source,
               const vertex_t& destination) -> int {
  // source must be within the bounds
  assert(g.has(source));

  // destination must be within the bounds
  assert(g.has(destination));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we only iterate as many times as cells that we have
  for (int i = g.height() * g.width(); i > 0; --i) {
    // go through all of the vertices
    for (int v = g.height() * g.width() - 1; v >= 0; --v) {
      _check_vertex(g, distances, v);
    }
  }

  return distances[destination.second][destination.first];
}

auto bellman_ford(const graph& g, const vertex_t& source)
    -> std::vector<std::vector<int>> {
  // source must be within the bounds
  assert(g.has(source));

  // we need to initialize the distances
  std::vector<std::vector<int>> distances(
      g.height(), std::vector(g.width(), graph::unreachable()));

  // source destination denotes the beginning where the cost is 0
  auto [sx, sy] = source;
  distances[sy][sx] = 0;

  // now we only iterate as many times as cells that we have
  for (int i = g.height() * g.width(); i > 0; --i) {
    // go through all of the vertices
    for (int v = g.height() * g.width() - 1; v >= 0; --v) {
      _check_vertex(g, distances, v);
    }
  }

  // now we check for the negative loops
  for (int v = g.height() * g.width() - 1; v >= 0; --v) {
    if (_check_vertex(g, distances, v, true)) {
      std::cerr << "[Bellman-Ford] Found a negative loop\n";
      break;
    }
  }

  return distances;
}


:::tip

You might've noticed that I've been using abbreviation BF interchangeably for both Bellman-Ford and brute-force. If you think about the way Bellman-Ford algorithm works, you should realize that in the worst case it's updating the shortest path till there no shorter path exists, so in a sense, you could really consider it a brute-force algorithm.

:::


  1. Breadth-first search ↩︎

  2. Of course, there are some technicalities like keeping track of the visited vertices to not taint the shortest path by already visited vertices. ↩︎

  3. or at least you should, LOL ↩︎