I have a situation in which I need to find optimal split positions in an array based on some costs. The problem goes like this:
As input I have an array of events ordered by an integer timestamp and as output I want an array of indexes which split the input array into many parts. The output array needs to be optimal (more on this below).
struct e {
int Time;
// other values
}
Example Input: [e0, e1, e2, e3, e4, e5, ..., e10]
Example output: [0, 2, 6, 8] (the 0 at the start is always there)
Using the above examples I can use the split indices to partition the original array into 5 subarrays like so:
[ [], [e0, e1], [e2, e3, e4, e5], [e6, e7], [e8, e9, e10] ]
The cost of this example solution is the total cost of "distances" between the subarrays:
double distance(e[] arr1, e[] arr2) {
// return distance from arr1 to arr2, order matters so non-euclidean
}
total cost = distance([], [e0, e1]) + distance([e0, e1], [e2, e3, e4, e5]) + ...
At this point it is helpful to understand the actual problem.
The input array represents musical notes at some time (i.e. a MIDI file) and I want to split the MIDI file into optimal guitar fingerings. Hence each subarray of notes represents a chord (or a melody grouped together in a single fingering). The distance between two subarrays represents the difficulty of moving from one fingering pattern to another. The goal is to find the easiest (optimal) way to play a song on a guitar.
I have not yet proved it but to me this looks like an NP-Complete or NP-Hard problem. Therefore it could be helpful if I could reduce this to another known problem and use a known divide and conquer algorithm. Also, one could solve this with a more traditional search algorithm (A* ?). It could be efficient because we can filter out bad solutions much faster than in a regular graph (because the input is technically a complete graph since each fingering can be reached from any other fingering).
I'm not able to decide what the best approach would be so I am currently stuck. Any tips or ideas would be appreciated.
It's probably not NP-hard.
Form a graph whose nodes correspond one-to-one to (contiguous) subarrays. For each pair of nodes u, v where u's right boundary is v's left, add an arc from u to v whose length is determined by distance(). Create an artificial source with an outgoing arc to each node whose left boundary is the beginning. Create an artificial sink with an incoming arc from each node whose right boundary is the end.
Now we can find a shortest path from the source to the sink via the linear-time (in the size of the graph, so cubic in the parameter of interest) algorithm for directed acyclic graphs.
This is a bit late but I did solve this problem. I ended up using a slightly modified version of Dijkstra for this but any pathfinding algo could work. I tried A* as well but finding a good heuristic proved to be extremely difficult because of the non-euclidean nature of the problem.
The main changes to Dijkstra are that at some point I can already tell that some unvisited nodes cannot provide an optimal result. This speeds up the algorithm by a lot which is also one of the reasons I didn't opt for A*.
The algorithm essentially works like this:
search()
visited = set()
costs = map<node, double>()
// add initial node to costs
while costs is not empty:
node = node with minimum cost in costs
if current.Index == songEnd:
// backtrack from current to get fingering for the song
return solution
visited.add(node)
foreach neighbour of node:
if visited node:
continue
newCost = costs[node] + distance(node, neighbour)
add neighbour with newCost to costs
// we can remove nodes that have a higher cost but
// which have a lower index than our current node
// this is because every fingering position is reachable
// from any fingering positions
// therefore a higher cost node which is not as far as our current node
// cannot provide an optimal solution
remove unoptimal nodes from costs
remove node from costs
// if costs ends up empty then it is impossible to play this song
// on guitar (e.g. more than 6 notes played at the same time)
The magic of this algorithm happens in fetching the neighbours and calculating the distance between two nodes but those are irrelevant for this question.
There is problem, I reduce it to a question as below:
In a connected undirected graph, edge weight is the time to go from one end to another. some people stand on some vertex. Now, they want to meet together, find a place(vertice) that within certain time T, all the people will arrive this assembly point. Try to minimise this T.
More information if you need for margin cases: No negative edge; cycle may exist; More than one person can stay on the same vertice; vertice may have no person; undirected edge, weight measures both u->v or v->u; people start from their initial location;
How to efficiently find it? Should I for every node v, calculate max(SPD(ui, v)) where ui are other people's locations, then choose the minimum one among these max times? Is there a better way?
I believe it could be done within a polynomial runtime bound as follows. In a first pass solve the All-Pairs Shortest Path problem to obtain a matrix with corresponding lengths of shortest paths for all vertices; afterwards iterate over the rows (or columns) and select a column where the maximum entry of all indices on which users are located.
It can be done by making parallel Dijkstra from all vertices, and stopping when sets of visited nodes intersect in one node. Intersection can be checked by counting. Algorithm sketch:
node_count = [1, 1, ...] * number_of_nodes # Number of visited sets node is in
dijkstras = set of objects D_n performing Dijsktra's algorithm starting from node n
queue = priority queue that stores tuples (first_in_queue_n, D_n).
first_in_queue_n is next node that will be visited by D_n
initialized by D_n.first_in_queue()
while:
first_in_queue_n, D_n = queue.pop_min()
node_count[first_in_queue_n] += 1
if node_count[first_in_queue_n] == number_of_nodes:
return first_in_queue_n
D_n.visite_node(first_in_queue_n)
queue.add( D_n.first_in_queue() )
I'm trying to find an efficient algorithm to generate a simple connected graph with given the number of nodes. Something like:
Input:
N - size of generated graph
Output:
simple connected graph G(v,e) with N vertices and S edges, The number of edges should be uniform distribution.
You might want to create a minimal spanning tree first to ensure connectivity. Later, randomly generate two nodes (which are not yet connected) and connect them. Repeat until you have S edges.
For minimal spanning tree, the easiest thing to do is start with a random node as tree. For every remaining node (ordered randomly), connect it to any node in the tree. The way you select the node within the tree (to connect to) defines the distribution of edges/node.
It may be a bit surprising but if you choose (randomly) log(n) edges (where n - maximum number of edges) then you can be almost sure that your graph is connected (a reference). If number of edges is much lower than log(n) you can be almost sure that the graph is disconnected.
How to generate the graph:
GenerateGraph(N,S):
if (log(N)/N) > S) return null // or you can take another action
V <- {0,...,N-1}
E <- {}
possible_edges <- {{v,u}: v,u in V} // all possible edges
while (size(E) < S):
e <- get_random_edge(possible_edges)
E.add(e)
possible_edges.remove(e)
G=(V,E)
if (G is connected)
return G
else
return GenerateGraph(N,S)
Let me know if you need any more guidance.
(Btw. right now I am dealing with exactly the same problem! Let me know if you need to generate some more sophisticated graphs:-))
A very common random graph generation algorithm (used in many academic works) is based on RMat method, or a generalization in the form of Kronecker graphs. It's a simple, iterative process, that uses very few parameters and is easily scalable.
There's a really good explanation here (including why it's better than other methods) -
http://www.graphanalysis.org/SIAM-PP08/Leskovic.pdf
There are versions of both implemented in many graph benchmark suites, for e.g.
BFS benchmark with a subdir with rMat generator - http://www.cs.cmu.edu/~pbbs/benchmarks/breadthFirstSearch.tar
Kronecker graph generator (both c and matlab codes included) - http://www.graph500.org/sites/default/files/files/graph500-2.1.4.tar.bz2
Given an undirected graph, I want to generate all subgraphs which are trees of size N, where size refers to the number of edges in the tree.
I am aware that there are a lot of them (exponentially many at least for graphs with constant connectivity) - but that's fine, as I believe the number of nodes and edges makes this tractable for at least smallish values of N (say 10 or less).
The algorithm should be memory-efficient - that is, it shouldn't need to have all graphs or some large subset of them in memory at once, since this is likely to exceed available memory even for relatively small graphs. So something like DFS is desirable.
Here's what I'm thinking, in pseudo-code, given the starting graph graph and desired length N:
Pick any arbitrary node, root as a starting point and call alltrees(graph, N, root)
alltrees(graph, N, root)
given that node root has degree M, find all M-tuples with integer, non-negative values whose values sum to N (for example, for 3 children and N=2, you have (0,0,2), (0,2,0), (2,0,0), (0,1,1), (1,0,1), (1,1,0), I think)
for each tuple (X1, X2, ... XM) above
create a subgraph "current" initially empty
for each integer Xi in X1...XM (the current tuple)
if Xi is nonzero
add edge i incident on root to the current tree
add alltrees(graph with root removed, N-1, node adjacent to root along edge i)
add the current tree to the set of all trees
return the set of all trees
This finds only trees containing the chosen initial root, so now remove this node and call alltrees(graph with root removed, N, new arbitrarily chosen root), and repeat until the size of the remaining graph < N (since no trees of the required size will exist).
I forgot also that each visited node (each root for some call of alltrees) needs to be marked, and the set of children considered above should only be the adjacent unmarked children. I guess we need to account for the case where no unmarked children exist, yet depth > 0, this means that this "branch" failed to reach the required depth, and cannot form part of the solution set (so the whole inner loop associated with that tuple can be aborted).
So will this work? Any major flaws? Any simpler/known/canonical way to do this?
One issue with the algorithm outlined above is that it doesn't satisfy the memory-efficient requirement, as the recursion will hold large sets of trees in memory.
This needs an amount of memory that is proportional to what is required to store the graph. It will return every subgraph that is a tree of the desired size exactly once.
Keep in mind that I just typed it into here. There could be bugs. But the idea is that you walk the nodes one at a time, for each node searching for all trees that include that node, but none of the nodes that were searched previously. (Because those have already been exhausted.) That inner search is done recursively by listing edges to nodes in the tree, and for each edge deciding whether or not to include it in your tree. (If it would make a cycle, or add an exhausted node, then you can't include that edge.) If you include it your tree then the used nodes grow, and you have new possible edges to add to your search.
To reduce memory use, the edges that are left to look at is manipulated in place by all of the levels of the recursive call rather than the more obvious approach of duplicating that data at each level. If that list was copied, your total memory usage would get up to the size of the tree times the number of edges in the graph.
def find_all_trees(graph, tree_length):
exhausted_node = set([])
used_node = set([])
used_edge = set([])
current_edge_groups = []
def finish_all_trees(remaining_length, edge_group, edge_position):
while edge_group < len(current_edge_groups):
edges = current_edge_groups[edge_group]
while edge_position < len(edges):
edge = edges[edge_position]
edge_position += 1
(node1, node2) = nodes(edge)
if node1 in exhausted_node or node2 in exhausted_node:
continue
node = node1
if node1 in used_node:
if node2 in used_node:
continue
else:
node = node2
used_node.add(node)
used_edge.add(edge)
edge_groups.append(neighbors(graph, node))
if 1 == remaining_length:
yield build_tree(graph, used_node, used_edge)
else:
for tree in finish_all_trees(remaining_length -1
, edge_group, edge_position):
yield tree
edge_groups.pop()
used_edge.delete(edge)
used_node.delete(node)
edge_position = 0
edge_group += 1
for node in all_nodes(graph):
used_node.add(node)
edge_groups.append(neighbors(graph, node))
for tree in finish_all_trees(tree_length, 0, 0):
yield tree
edge_groups.pop()
used_node.delete(node)
exhausted_node.add(node)
Assuming you can destroy the original graph or make a destroyable copy I came up to something that could work but could be utter sadomaso because I did not calculate its O-Ntiness. It probably would work for small subtrees.
do it in steps, at each step:
sort the graph nodes so you get a list of nodes sorted by number of adjacent edges ASC
process all nodes with the same number of edges of the first one
remove those nodes
For an example for a graph of 6 nodes finding all size 2 subgraphs (sorry for my total lack of artistic expression):
Well the same would go for a bigger graph, but it should be done in more steps.
Assuming:
Z number of edges of most ramificated node
M desired subtree size
S number of steps
Ns number of nodes in step
assuming quicksort for sorting nodes
Worst case:
S*(Ns^2 + MNsZ)
Average case:
S*(NslogNs + MNs(Z/2))
Problem is: cannot calculate the real omicron because the nodes in each step will decrease depending how is the graph...
Solving the whole thing with this approach could be very time consuming on a graph with very connected nodes, however it could be paralelized, and you could do one or two steps, to remove dislocated nodes, extract all subgraphs, and then choose another approach on the remainder, but you would have removed a lot of nodes from the graph so it could decrease the remaining run time...
Unfortunately this approach would benefit the GPU not the CPU, since a LOT of nodes with the same number of edges would go in each step.... and if parallelization is not used this approach is probably bad...
Maybe an inverse would go better with the CPU, sort and proceed with nodes with the maximum number of edges... those will be probably less at start, but you will have more subgraphs to extract from each node...
Another possibility is to calculate the least occuring egde count in the graph and start with nodes that have it, that would alleviate the memory usage and iteration count for extracting subgraphs...
Unless I'm reading the question wrong people seem to be overcomplicating it.
This is just "all possible paths within N edges" and you're allowing cycles.
This, for two nodes: A, B and one edge your result would be:
AA, AB, BA, BB
For two nodes, two edges your result would be:
AAA, AAB, ABA, ABB, BAA, BAB, BBA, BBB
I would recurse into a for each and pass in a "template" tuple
N=edge count
TempTuple = Tuple_of_N_Items ' (01,02,03,...0n) (Could also be an ordered list!)
ListOfTuple_of_N_Items ' Paths (could also be an ordered list!)
edgeDepth = N
Method (Nodes, edgeDepth, TupleTemplate, ListOfTuples, EdgeTotal)
edgeDepth -=1
For Each Node In Nodes
if edgeDepth = 0 'Last Edge
ListOfTuples.Add New Tuple from TupleTemplate + Node ' (x,y,z,...,Node)
else
NewTupleTemplate = TupleTemplate + Node ' (x,y,z,Node,...,0n)
Method(Nodes, edgeDepth, NewTupleTemplate, ListOfTuples, EdgeTotal
next
This will create every possible combination of vertices for a given edge count
What's missing is the factory to generate tuples given an edge count.
You end up with a list of possible paths and the operation is Nodes^(N+1)
If you use ordered lists instead of tuples then you don't need to worry about a factory to create the objects.
If memory is the biggest problem you can use a NP-ish solution using tools from formal verification. I.e., guess a subset of nodes of size N and check whether it's a graph or not. To save space you can use a BDD (http://en.wikipedia.org/wiki/Binary_decision_diagram) to represent the original graph's nodes and edges. Plus you can use a symbolic algorithm to check if the graph you guessed is really a graph - so you don't need to construct the original graph (nor the N-sized graphs) at any point. Your memory consumption should be (in big-O) log(n) (where n is the size of the original graph) to store the original graph, and another log(N) to store every "small graph" you want.
Another tool (which is supposed to be even better) is to use a SAT solver. I.e., construct a SAT formula that is true iff the sub-graph is a graph and supply it to a SAT solver.
For a graph of Kn there are approximately n! paths between any two pairs of vertices. I haven't gone through your code but here is what I would do.
Select a pair of vertices.
Start from a vertex and try to reach the destination vertex recursively (something like dfs but not exactly). I think this would output all the paths between the chosen vertices.
You could do the above for all possible pairs of vertices to get all simple paths.
It seems that the following solution will work.
Go over all partitions into two parts of the set of all vertices. Then count the number of edges which endings lie in different parts (k); these edges correspond to the edge of the tree, they connect subtrees for the first and the second parts. Calculate the answer for both parts recursively (p1, p2). Then the answer for the entire graph can be calculated as sum over all such partitions of k*p1*p2. But all trees will be considered N times: once for each edge. So, the sum must be divided by N to get the answer.
Your solution as is doesn't work I think, although it can be made to work. The main problem is that the subproblems may produce overlapping trees so when you take the union of them you don't end up with a tree of size n. You can reject all solutions where there is an overlap, but you may end up doing a lot more work than needed.
Since you are ok with exponential runtime, and potentially writing 2^n trees out, having V.2^V algorithms is not not bad at all. So the simplest way of doing it would be to generate all possible subsets n nodes, and then test each one if it forms a tree. Since testing whether a subset of nodes form a tree can take O(E.V) time, we are potentially talking about V^2.V^n time, unless you have a graph with O(1) degree. This can be improved slightly by enumerating subsets in a way that two successive subsets differ in exactly one node being swapped. In that case, you just have to check if the new node is connected to any of the existing nodes, which can be done in time proportional to number of outgoing edges of new node by keeping a hash table of all existing nodes.
The next question is how do you enumerate all the subsets of a given size
such that no more than one element is swapped between succesive subsets. I'll leave that as an exercise for you to figure out :)
I think there is a good algorithm (with Perl implementation) at this site (look for TGE), but if you want to use it commercially you'll need to contact the author. The algorithm is similar to yours in the question but avoids the recursion explosion by making the procedure include a current working subtree as a parameter (rather than a single node). That way each edge emanating from the subtree can be selectively included/excluded, and recurse on the expanded tree (with the new edge) and/or reduced graph (without the edge).
This sort of approach is typical of graph enumeration algorithms -- you usually need to keep track of a handful of building blocks that are themselves graphs; if you try to only deal with nodes and edges it becomes intractable.
This algorithm is big and not easy one to post here. But here is link to reservation search algorithm using which you can do what you want. This pdf file contains both algorithms. Also if you understand russian you can take a look to this.
So you have a graph with with edges e_1, e_2, ..., e_E.
If I understand correctly, you are looking to enumerate all subgraphs which are trees and contain N edges.
A simple solution is to generate each of the E choose N subgraphs and check if they are trees.
Have you considered this approach? Of course if E is too large then this is not viable.
EDIT:
We can also use the fact that a tree is a combination of trees, i.e. that each tree of size N can be "grown" by adding an edge to a tree of size N-1. Let E be the set of edges in the graph. An algorithm could then go something like this.
T = E
n = 1
while n<N
newT = empty set
for each tree t in T
for each edge e in E
if t+e is a tree of size n+1 which is not yet in newT
add t+e to newT
T = newT
n = n+1
At the end of this algorithm, T is the set of all subtrees of size N. If space is an issue, don't keep a full list of the trees, but use a compact representation, for instance implement T as a decision tree using ID3.
I think problem is under-specified. You mentioned that graph is undirected and that subgraph you are trying to find is of size N. What is missing is number of edges and whenever trees you are looking for binary or you allowed to have multi-trees. Also - are you interested in mirrored reflections of same tree, or in other words does order in which siblings are listed matters at all?
If single node in a tree you trying to find allowed to have more than 2 siblings which should be allowed given that you don't specify any restriction on initial graph and you mentioned that resulting subgraph should contain all nodes.
You can enumerate all subgraphs that have form of tree by performing depth-first traversal. You need to repeat traversal of the graph for every sibling during traversal. When you'll need to repeat operation for every node as a root.
Discarding symmetric trees you will end up with
N^(N-2)
trees if your graph is fully connected mesh or you need to apply Kirchhoff's Matrix-tree theorem