Dijktra’s Algorithm and Single-Source Shortest Paths. Suppose we want to record the shortest paths from some source to every single other vertex (so that we can rapidly found a route from s to X, from s to Y, and so forth). We already know how to do this if we’re only counting the number of edges, we just use BFS.
But if edges have weights (representing, for example road lengths), we have to do something else. It turns out that even considering edge weights, we can preprocess the shortest route from the source to every vertex very efficiently. We store the answer as a “shortest paths tree”. Typically, a shortest paths tree is stored as an array of edgeTo values (and optionally distTo values if we want a constant time distTo() operation).
To find the SPT, we can use Dijkstra’s algorithm, which is quite simple once you understand it. Essentially, we visit each vertex in order of its distance from the source, where each visit consists of relaxing every edge. Informally, relaxing an edge means using it if its better than the best known distance to the target vertex, otherwise ignoring it. Or in pseudocode:
Dijkstra(G, s): while not every vertex has been visited: visit(unmarked vertex v for which distTo(v) is minimized)
Where visit is given by the following pseudocode:
visit(v): mark(v) for each edge e of s: relax(e)
And finally, relax is given by:
relax(e): v = e.source w = e.target currentBestKnownWeight = distTo(w) possiblyBetterWeight = distTo(v) + e.weight if possiblyBetterWeight < currentBestKnownWeight Use e instead of whatever we were using before
Runtime is $O(V \times \log V + V \times \log V + E \times \log V)$, and since $E \gt V$ for any graph we’d run Dijkstra’s algorithm on, this can be written as more simply O(E log V). See slides for runtime description.
A* Single-Target Shortest Paths. If we need only the path to a single target, then Dijkstra’s is inefficient as it explores many many edges that we don’t care about (e.g. when routing from Denver to NYC, we’d explore everything within more than a thousand miles in all directions before reaching NYC).
To fix this, we make a very minor change to Dijkstra’s, where instead of visiting vertices in order of distance from the source, we visit them in order of distance from the source + h(v), where h(v) is some heuristic.
Or in pseudocode:
A*(G, s): while not every vertex has been visited: visit(unmarked vertex v for which distTo(v) + h(v) is minimized)
It turns out (but we did not prove), that as long as h(v) is less than the true distance from s to v, then the result of A* will always be correct.
Note: In the version in class, we did not use an explicit ‘mark’. Instead, we tossed everything in the PQ, and we effectively considered a vertex marked if it had been removed from the PQ.
Suppose we have an edgeTo and a distTo array, how would we implement a
public int distTo(int w)method? How would we implement a
public Iterable<int> pathTo(int w)method, where
distToreturns the length of the shortest path and
pathToreturns an Iterable of vertices where the 0th element is the source, the 1st element in the next item from the source towards w, and so forth.
Problem 4 from Princeton’s Fall 2009 final.
True or false: Adding a constant to every edge weight does not change the solution to the single-source shortest-paths problem.
True or false: Multiplying a positive constant to every edge weight does not change the solution to the single-source shortest-paths problem.
(great problem) Problem 4 from Princeton’s Fall 2011 final. Note that when the exam says to “relax a vertex”, that mean to relax all of a vertex’s edges. A similar problem is given as #6 on Princeton’s Spring 2012 final.
Problem 5 from Princeton’s Fall 2012 final.
Problem 1 from Kartik’s mock final
Adapted from Algorithms 4.4.25: Given a digraph with positive edge weights, and two distinguished subsets of vertices S and T, find a shortest path from any vertex in S to any vertex in T. Your algorithm should run in time proportional to E log V, in the worst case.
Problem 5 from my Spring 2013 final.
Problem 6 from Kartik’s Algorithm Worksheet
Describe a family of graphs with V vertices and E edges for which the worst-case running time of Dijkstra’s algorithm is achieved.
Iterative DFS: Problem 6 from this semester’s discussion worksheet provides a flawed implementation of DFS.
Adapted from Algorithms 4.4.34. Give an algorithm to solve the following problem: Given a weighted digraph, find a monotonic shortest path from s to every other vertex. A path is monotonic if the weight of every edge on the path is either strictly increasing or strictly decreasing. The path should be simple (no repeated vertices).
Adapted from Algorithms 4.4.37. Develop an algorithm for finding an edge whose removal causes maximal increase in the shortest-paths length from one given vertex to another given vertex in a given edge-weighted digraph.
- Problem 12 from my http://www.cs.princeton.edu/courses/archive/fall13/cos226/exams/fin-f13.pdf.