We have successfully expressed 1-QMCP as a min cost circulation problem. This solution leads to push-relabel + cost-scaling algorithm implementation. The push-relabel + cost-scaling CUDA solution may be problematic, while there are papers on push-relabel algorithms that claim to speed up CPU solutions by 3-8 times (here), there is not many about the cost-scaling variant of the algorithm.
The cost-scaling seems to be compatible with the push-relabel CUDA solution, but it brings risks, since we don't now much about the performance.
To do
We would like to express 1-QMCP as a linear programming problem. It's important to find the direct solution, the solutions like this: 1-qmcp -> min cost max flow -> linear programming are not valid, since they increase complexity and probably slow down the algorithm, we need 1-qmcp -> linear programming.
The best place to start is here (2. Transformation of matrices with consecutive 1s)
Background:
We have successfully expressed 1-QMCP as a min cost circulation problem. This solution leads to push-relabel + cost-scaling algorithm implementation. The push-relabel + cost-scaling CUDA solution may be problematic, while there are papers on push-relabel algorithms that claim to speed up CPU solutions by 3-8 times (here), there is not many about the cost-scaling variant of the algorithm.
The cost-scaling seems to be compatible with the push-relabel CUDA solution, but it brings risks, since we don't now much about the performance.
To do
We would like to express 1-QMCP as a linear programming problem. It's important to find the direct solution, the solutions like this:
1-qmcp -> min cost max flow -> linear programmingare not valid, since they increase complexity and probably slow down the algorithm, we need1-qmcp -> linear programming.The best place to start is here (2. Transformation of matrices with consecutive 1s)