two sided dijkstra #268
two sided dijkstra #268
Conversation
|
I added a very simple test, not sure if it is sufficient |
|
Tests are failing because of #271 which has now been merged. If we merge the |
src/shortestpaths/dijkstra.jl
Outdated
| end | ||
| # keep weight of the best seen path and the midpoint vertex | ||
| mu, mid_v = typemax(T), -1 | ||
|
|
There was a problem hiding this comment.
All that white space may not be necessary
| uf, ub = dequeue!(Qf), dequeue!(Qb) | ||
|
|
||
| for v in outneighbors(g, uf) | ||
| relax(uf, v, distmx, dists_f, parents_f, visited_f, Qf) |
There was a problem hiding this comment.
This relax function clarifies things, would it be hard to add it to the original dijkstra too?
There was a problem hiding this comment.
I was hoping you would notice =p I'll make the changes
There was a problem hiding this comment.
I have a new version of the relax function that should work for both the dijkstra_shortest_paths and bidijkstra_shortest_path but it is a bit clunky. I am wondering; why the need for the parents and preds structures ? wouldn't the list of predecessors be enough ?
A cosmetic note: should we rename the function the bidijkstra_shortest_paths if it should be able to deal with multiple paths ?
src/shortestpaths/dijkstra.jl
Outdated
| end | ||
| ds_f = DijkstraState{T,U}(parents_f, dists_f, preds_f, zeros(nvg), Vector{U}()) | ||
| ds_b = DijkstraState{T,U}(parents_b, dists_b, preds_b, zeros(nvg), Vector{U}()) | ||
| path = vcat(enumerate_paths(ds_f, mid_v), reverse(enumerate_paths(ds_b, mid_v)[1:end-1])) |
There was a problem hiding this comment.
I'm not sure how to deal with such situations, as far as I can tell multiple paths with the same cost can be detected if the condition mu == ls + lt is reached. I'll look into it
There was a problem hiding this comment.
Given that bidijkstra_shortest_path only returns src-dst-paths it doesn't make much sense to keep a full list of predecessors, does it ? I am not sure what the most elegant way to deal with these situations, especially when trying to share code with dijkstra_shortest_paths.
I've also noticed that there doesn't seem to be a straightforward way to retrieve multiple paths for a single (src,dst) pair. As far a I can tell the enumerate_paths routine returns a single path per destination. Should this be addressed at some point ?
|
Not sure why tests are failing |
|
I think your modification to Dijkstra somehow broke a test in the centrality part of the package.
The error comes from the following test: # test #1405 / #1406
g = GenericGraph(grid([50, 50]))
z = betweenness_centrality(g; normalize=false)
@test maximum(z) < nv(g) * (nv(g) - 1)and it looks like this: It is related to the following issue in the old LightGraphs.jl package |
There was a problem hiding this comment.
There seems to be some different behavior for the pathcount when using dijkstra_shortest_path. Might be some kind of overflow issue, as I only saw it on large graphs so far.
before:
julia> dijkstra_shortest_paths(grid([50, 50]), 1).pathcounts[50*50]
2.5477612258980867e28after:
julia> dijkstra_shortest_paths(grid([50, 50]), 1).pathcounts[50*50]
8.581105107791177e17I think the right solution here should be (2 * 49 \choose 49), i.e
julia> binomial(BigInt(2*49), 49)
25477612258980856902730428600so apart from some rounding errors, the old solution looks much more correct to me.
src/shortestpaths/dijkstra.jl
Outdated
| @@ -83,7 +83,7 @@ function dijkstra_shortest_paths( | |||
| parents = zeros(U, nvg) | |||
| visited = zeros(Bool, nvg) | |||
|
|
|||
| pathcounts = zeros(nvg) | |||
There was a problem hiding this comment.
We probably should not change this, it might overflow.
|
It also look to me, as if those changes might make the algorithm perform slower than before, but I only checked this single case: julia> @btime dijkstra_shortest_paths($(grid([50, 50])), 1);
548.740 μs (81 allocations: 165.11 KiB)after julia> @btime dijkstra_shortest_paths($(grid([50, 50])), 1);
585.437 μs (83 allocations: 184.72 KiB) |
…destination of the shortest path
…on), tests now work
|
@Dolgalad is this ready for another round of review? sorry for the delay |
Implemented the Two sided Dijkstra algorithm, still requires documentation and tests