I am solving an optimization problem in which, among other things, I must maximize flow networks. I implemented a c++ code based flow-maximization algorithm based in the following java code that appears in the book of Sedgewick "Algorithms in Java, Third Edition, Part 5: Graph Algorithms", which maximizes a network flow using a vertex-based PREFLOW-push algorithm:
class NetworkMaxFlow
{ private Network G; private int s, t;
private int[] h, wt;
private void initheights()
NetworkMaxFlow(Network G, int s, int t)
{ this.G = G; this.s = s; this.t = t;
wt = new int[G.V()]; h = new int[G.V()];
initheights();
intGQ gQ = new intGQ(G.V());
gQ.put(s); wt[t] = -(wt[s] = Edge.M*G.V());
while (!gQ.empty())
{ int v = gQ.get();
AdjList A = G.getAdjList(v);
for (Edge e = A.beg(); !A.end(); e = A.nxt())
{ int w = e.other(v), cap = e.capRto(w);
int P = cap < wt[v] ? cap : wt[v];
if (P > 0 && v == s || h[v] == h[w]+1) // first observation (see below)
{ e.addflowRto(w, P);
wt[v] -= P; wt[w] += P;
if ((w != s) && (w != t)) gQ.put(w); // enqueue w if it is not source or sink
}
}
if (v != s && v != t && wt[v] > 0) // why check v != t if t never enter the queue?
{ h[v]++; gQ.put(v); }
}
}
}
My implementation, based on that code, fails to maximize the following network After execution, the the resulting flow is as follows With this flow, the flow value is 8, but the maximum is 9, as indicated by the flow of the figure below According to my understanding, the algorithm is consistent with the explanation of the book. However, I see two strange things
- There is no explicit preflow phase from source. It is included in the
while
and executed first and only once when the predicateP > 0 && v == s
is true. Maybe this was done to shorten the code - According to my understanding, and discourse of the book, the sink never enters the queue. However, when the height is increased, the code checks that v != t. Any reason for this?
This is an excerpt from my implementation of this algorithm in C++
template <class Net, class Q_Type> typename Net::Flow_Type
generic_preflow_vertex_push_maximum_flow(Net & net)
{
init_height_in_nodes(net); // breadth first traverse from sink to
// source. Nodes are labeled with their
// minimal distance (in nodes) to sink
auto source = net.get_source();
auto sink = net.get_sink();
using Itor = __Net_Iterator<Net>;
Q_Type q; // generic queue (can be fifo, heap or random) of active nodes
// preflow: floods all nodes connected to the source
for (Itor it(source); it.has_curr(); it.next())
{
auto arc = it.get_curr();
arc->flow = arc->cap; // saturate arc to its maximum
auto tgt = net.get_tgt_node(arc);
put_in_active_queue(q, tgt);
assert(node_height<Net>(source) == node_height<Net>(tgt) + 1);
assert(not is_residual<Net>(source, arc));
}
while (not q.is_empty()) // while there are active nodes
{
auto src = get_from_active_queue(q);
auto excess = net.get_in_flow(src) - net.get_out_flow(src);
for (Itor it(src); it.has_curr(); it.next())
{
auto arc = it.get_curr();
auto tgt = net.get_connected_node(arc, src);
if (node_height<Net>(src) != node_height<Net>(tgt) + 1)
continue; // this arc is not eligible
typename Net::Flow_Type flow_to_push;
if (is_residual<Net>(src, arc))
{
flow_to_push = std::min(arc->flow, excess);
arc->flow -= flow_to_push;
}
else
{
flow_to_push = std::min(arc->cap - arc->flow, excess);
arc->flow += flow_to_push;
}
excess -= flow_to_push;
if (tgt != sink and tgt != source)
put_in_active_queue(q, tgt);
}
if (excess > 0) // src still active?
{
node_height<Net>(src)++;
put_in_active_queue(q, src);
}
}
return net.flow_value(); // sum of all outing flow from source
}
¿Someone find any logical inconsistency between my code and the code of Sedgewick? I have the impression that my code (and perhaps also the Sedgewick) is not properly handling the increases in heights. But I do not manage to understand why
I show a detailed execution trace with the network that fails to maximize (the trace start from the first q.get() of while. The values in parentheses are the values of the heights. IN is the incoming flow to the node. OUT the outcoming one.
As example, the line
4104 (2) --> 0 (1) pushing 1 from 4104 toward 0
refers the eligible arc 4104-->0. The node 4104 has height 2 and the node 0 has height 1. The expression "pushing 1" means that 1 unity of flow is pushed toward the target node (0). The line ================
separates each queue extraction. The queue is FIFO and its state is printed at the end of each processing.
Note that many times zero flow units are pushed or reduceed, but the destination node becomes active.
This is the execution trace
Initial Queue = 4104 4105 4106 4107 4108
Active node 4104 Height = 2 IN = 1 OUT = 0
4104 (2) --> source (3) not eligible
4104 (2) --> 0 (1) pushing 1 from 4104 toward 0
4104 (2) --> 1 (1) pushing 0 from 4104 toward 1
4104 (2) --> 2 (1) pushing 0 from 4104 toward 2
4104 (2) --> 4 (1) pushing 0 from 4104 toward 4
Excess = 0
Queue = 4105 4106 4107 4108 0 1 2 4
================
Active node 4105 Height = 2 IN = 3 OUT = 0
4105 (2) --> source (3) not eligible
4105 (2) --> 1 (1) pushing 1 from 4105 toward 1
4105 (2) --> 4 (1) pushing 1 from 4105 toward 4
4105 (2) --> 6 (1) pushing 1 from 4105 toward 6
Excess = 0
Queue = 4106 4107 4108 0 1 2 4 6
================
Active node 4106 Height = 2 IN = 1 OUT = 0
4106 (2) --> source (3) not eligible
4106 (2) --> 1 (1) pushing 1 from 4106 toward 1
4106 (2) --> 5 (1) pushing 0 from 4106 toward 5
Excess = 0
Queue = 4107 4108 0 1 2 4 6 5
================
Active node 4107 Height = 2 IN = 1 OUT = 0
4107 (2) --> source (3) not eligible
4107 (2) --> 1 (1) pushing 1 from 4107 toward 1
4107 (2) --> 2 (1) pushing 0 from 4107 toward 2
4107 (2) --> 3 (1) pushing 0 from 4107 toward 3
4107 (2) --> 4 (1) pushing 0 from 4107 toward 4
4107 (2) --> 6 (1) pushing 0 from 4107 toward 6
Excess = 0
Queue = 4108 0 1 2 4 6 5 3
================
Active node 4108 Height = 2 IN = 3 OUT = 0
4108 (2) --> source (3) not eligible
4108 (2) --> 1 (1) pushing 1 from 4108 toward 1
4108 (2) --> 2 (1) pushing 1 from 4108 toward 2
4108 (2) --> 4 (1) pushing 1 from 4108 toward 4
4108 (2) --> 5 (1) pushing 0 from 4108 toward 5
4108 (2) --> 6 (1) pushing 0 from 4108 toward 6
Excess = 0
Queue = 0 1 2 4 6 5 3
================
Active node 0 Height = 1 IN = 1 OUT = 0
0 (1) --> sink (0) pushing 1 from 0 toward sink
0 (1) --> 4104 (2) not eligible
Excess = 0
Queue = 1 2 4 6 5 3
================
Active node 1 Height = 1 IN = 4 OUT = 0
1 (1) --> sink (0) pushing 2 from 1 toward sink
1 (1) --> 4105 (2) not eligible
1 (1) --> 4106 (2) not eligible
1 (1) --> 4107 (2) not eligible
1 (1) --> 4108 (2) not eligible
Excess = 2 1 goes back onto queue with label 2
Queue = 2 4 6 5 3 1
================
Active node 2 Height = 1 IN = 1 OUT = 0
2 (1) --> sink (0) pushing 1 from 2 toward sink
2 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 4 6 5 3 1
================
Active node 4 Height = 1 IN = 2 OUT = 0
4 (1) --> sink (0) pushing 2 from 4 toward sink
4 (1) --> 4105 (2) not eligible
4 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 6 5 3 1
================
Active node 6 Height = 1 IN = 1 OUT = 0
6 (1) --> sink (0) pushing 1 from 6 toward sink
6 (1) --> 4105 (2) not eligible
Excess = 0
Queue = 5 3 1
================
Active node 5 Height = 1 IN = 0 OUT = 0
5 (1) --> sink (0) pushing 0 from 5 toward sink
Excess = 0
Queue = 3 1
================
Active node 3 Height = 1 IN = 0 OUT = 0
3 (1) --> sink (0) pushing 0 from 3 toward sink
Excess = 0
Queue = 1
================
Active node 1 Height = 2 IN = 4 OUT = 2
1 (2) --> 4105 (2) not eligible
1 (2) --> 4106 (2) not eligible
1 (2) --> 4107 (2) not eligible
1 (2) --> 4108 (2) not eligible
Excess = 2 1 goes back onto queue with label 3
Queue = 1
================
Active node 1 Height = 3 IN = 4 OUT = 2
1 (3) --> 4105 (2) Reducing 1 from 1 toward 4105
1 (3) --> 4106 (2) Reducing 1 from 1 toward 4106
1 (3) --> 4107 (2) Reducing 0 from 1 toward 4107
1 (3) --> 4108 (2) Reducing 0 from 1 toward 4108
Excess = 0
Queue = 4105 4106 4107 4108
================
Active node 4105 Height = 2 IN = 3 OUT = 2
4105 (2) --> source (3) not eligible
4105 (2) --> 1 (3) not eligible
Excess = 1 4105 goes back onto queue with label 3
Queue = 4106 4107 4108 4105
================
Active node 4106 Height = 2 IN = 1 OUT = 0
4106 (2) --> source (3) not eligible
4106 (2) --> 1 (3) not eligible
4106 (2) --> 5 (1) pushing 1 from 4106 toward 5
Excess = 0
Queue = 4107 4108 4105 5
================
Active node 4107 Height = 2 IN = 1 OUT = 1
4107 (2) --> source (3) not eligible
4107 (2) --> 2 (1) pushing 0 from 4107 toward 2
4107 (2) --> 3 (1) pushing 0 from 4107 toward 3
4107 (2) --> 4 (1) pushing 0 from 4107 toward 4
4107 (2) --> 6 (1) pushing 0 from 4107 toward 6
Excess = 0
Queue = 4108 4105 5 2 3 4 6
================
Active node 4108 Height = 2 IN = 3 OUT = 3
4108 (2) --> source (3) not eligible
4108 (2) --> 5 (1) pushing 0 from 4108 toward 5
4108 (2) --> 6 (1) pushing 0 from 4108 toward 6
Excess = 0
Queue = 4105 5 2 3 4 6
================
Active node 4105 Height = 3 IN = 3 OUT = 2
4105 (3) --> source (3) not eligible
4105 (3) --> 1 (3) not eligible
Excess = 1 4105 goes back onto queue with label 4
Queue = 5 2 3 4 6 4105
================
Active node 5 Height = 1 IN = 1 OUT = 0
5 (1) --> sink (0) pushing 1 from 5 toward sink
5 (1) --> 4106 (2) not eligible
Excess = 0
Queue = 2 3 4 6 4105
================
Active node 2 Height = 1 IN = 1 OUT = 1
2 (1) --> sink (0) pushing 0 from 2 toward sink
2 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 3 4 6 4105
================
Active node 3 Height = 1 IN = 0 OUT = 0
3 (1) --> sink (0) pushing 0 from 3 toward sink
Excess = 0
Queue = 4 6 4105
================
Active node 4 Height = 1 IN = 2 OUT = 2
4 (1) --> 4105 (4) not eligible
4 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 6 4105
================
Active node 6 Height = 1 IN = 1 OUT = 1
6 (1) --> sink (0) pushing 0 from 6 toward sink
6 (1) --> 4105 (4) not eligible
Excess = 0
Queue = 4105
================
Active node 4105 Height = 4 IN = 3 OUT = 2
4105 (4) --> source (3) Reducing 1 from 4105 toward source
4105 (4) --> 1 (3) pushing 0 from 4105 toward 1
Excess = 0
Queue = 1
================
Active node 1 Height = 3 IN = 2 OUT = 2
1 (3) --> 4107 (2) Reducing 0 from 1 toward 4107
1 (3) --> 4108 (2) Reducing 0 from 1 toward 4108
Excess = 0
Queue = 4107 4108
================
Active node 4107 Height = 2 IN = 1 OUT = 1
4107 (2) --> source (3) not eligible
4107 (2) --> 2 (1) pushing 0 from 4107 toward 2
4107 (2) --> 3 (1) pushing 0 from 4107 toward 3
4107 (2) --> 4 (1) pushing 0 from 4107 toward 4
4107 (2) --> 6 (1) pushing 0 from 4107 toward 6
Excess = 0
Queue = 4108 2 3 4 6
================
Active node 4108 Height = 2 IN = 3 OUT = 3
4108 (2) --> source (3) not eligible
4108 (2) --> 5 (1) pushing 0 from 4108 toward 5
4108 (2) --> 6 (1) pushing 0 from 4108 toward 6
Excess = 0
Queue = 2 3 4 6 5
================
Active node 2 Height = 1 IN = 1 OUT = 1
2 (1) --> sink (0) pushing 0 from 2 toward sink
2 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 3 4 6 5
================
Active node 3 Height = 1 IN = 0 OUT = 0
3 (1) --> sink (0) pushing 0 from 3 toward sink
Excess = 0
Queue = 4 6 5
================
Active node 4 Height = 1 IN = 2 OUT = 2
4 (1) --> 4105 (4) not eligible
4 (1) --> 4108 (2) not eligible
Excess = 0
Queue = 6 5
================
Active node 6 Height = 1 IN = 1 OUT = 1
6 (1) --> sink (0) pushing 0 from 6 toward sink
6 (1) --> 4105 (4) not eligible
Excess = 0
Queue = 5
================
Active node 5 Height = 1 IN = 1 OUT = 1
5 (1) --> sink (0) pushing 0 from 5 toward sink
5 (1) --> 4106 (2) not eligible
Excess = 0
Queue =
Your questions
Preflow
Yes, I think that was done to compact the code. Due to the assignment
wt[s] = Edge.M*G.V()
at the beginning the source has enough "virtual" excess in order to fuel the preflow in the first iteration of the algorithm. If you wanted to do the same trick in your implementation, you would have to inflate theexcess
variable (that you calculate on the fly instead of storing it in an array like Sedgewick) during the first iteration, when you encounter the source node. But you don't have to do it that way, your explicit preflooding seems just fine and probably even makes things more readable.I also suspect that in the condition
P > 0 && v == s || h[v] == h[w]+1
the implicit operator precedence(P > 0 && v == s) || h[v] == h[w]+1
was not intended. The way it is written, the checkP > 0
is only performed for the source case. Not checking it in the other cases won't hurt because executing the body withP == 0
will just push a 0 excess to other nodes (which doesn't do anything) and unnecessarily add those other nodes to the queue, just for them being removed again immediately. I've seen that you implemented it the same way. That's OK, even though a slight waste of computational time. ProbablyP > 0 && (v == s || h[v] == h[w]+1)
was really intended.Sink in the queue
I agree, this is weird. The only way the sink can enter the queue is by being the source at the same time (in a graph with only one node and no edges). However, in that case the condition
v != s
already avoids the infinite loop, no extra condition necessary.Wrong results
Initial heights
You are getting wrong results because your initial heights are wrong. I cannot compare Sedgewick's initialization code with yours because your post doesn't specify either. But I learn from your log file that you are starting with a height of 3 for
source
. That is clearly against the conditions for height functions: Thesource
has to start out with a height equal to the number of nodes in your graph and will keep it during the whole algorithm.In your case the height of
source
is too close to the height of its downstream neighbors. This causes the downstream neighbors to push some flow back to the source quite soon, before trying hard enough to let it flow further downstream. They are supposed to do that only if there is no way to let it flow down to the sink.Broken graph?
What further concerns me is that the edge
[4104 -> 1]
seems to disappear. While it is mentioned during the processing of the node 4104, it never shows up during the processing of the node 1. All other incoming edges are mentioned, be it with regard to "not eligible", "pushing" or "reducing". I'd expect it to appear as one of the three types, just like the others. But then again, I don't know where exactly you put those logging statements. Still, it leaves a bit of a concern and I thought I'd mention it in case you still have trouble after fixing your heights.