Improving Accuracy and Quality (anneal_schedule)
I'm trying to improve the accuracy and "quality" results of my QUBO problem. I tried the "quadratic growth", pause and quench from
But the only way to improve accuracy is increasing the number of "num_reads" in the: EmbeddingComposite(DWaveSampler().sample_qubo(bqm, num_reads=XXX, anneal_schedule=YYY).
e.g.: The Optimal value is: -500.03, but the Response (first 5) is:
sequence / energy
Accuracy: the best I can get after 10000 "num_reads" is (-442.28) / (-500.03) = 88% accuracy.
Quality: getting more results close to the minimal value (improving new results), but:
- Roughly 98% of the values are very far from the minimal value.
- Even after finding the minimal value (-442.28 in the sequence 4464) the next samples/sequences didn't improve it.
Seems that the samples are totally random, not exploring the local minimum and never finding the global minimum.
What can I do? Reverse Annealing? Some change in the Sampler or Composite? Changes in the Annealing Time?
Sometimes the problem formulation can be a bit tricky.
Since this is running on a physical machine, sometimes we run into physical limitations.
One thing that is problematic and can be difficult to resolve is maximizing the energy gap.
The energy gap is the energy between the ground state and the first excited state.
Here is a link to some information on fine tuning your problem:
In particular take a look at "Overcoming Imprecisions of Qubit Biases and Coupling Strengths" and "Controlling the Energy Gap".
This documentation should also be useful in understanding how performance can decline:
It describes various sources of error associated with the QPU.
Hopefully this was helpful. I will update if I get any further information that might be of use.
The links do not work. Could you please post valid links.
All of the above links should be up to date now.
Please sign in to leave a comment.