Chain Break Fraction

I executed some experiments using "sample_qubo" and "num_reads=1000". I realized that 80% of experiments have a "chain_break_fraction" between 1% and 22%. I wonder:

1) what is the maximum "chain_break_fraction" percentage to the "response.first.energy" be considered a valid answer? (zero?)

2) is there some way to remove/ignore, from the response, all the energies (num_reads) with a non-zero chain_break_fraction? (purging the response)

3) is there some way to find the best value for "chain_strength"? (only trying several combinations?)

4) what more can I do to improve the response quality?

Thank you so much.

0

Comments

1 comment
  • 1) Chain break fraction is not necessarily predictive of success in optimization. In the embedding tools majority vote is applied to broken chains by default, states with broken chains are mapped into states without broken chains, and these states can have good properties (disproportionately sample the optima).
    2) The majority voted (corrected) samples are not marked as such in the embedding solvers. If you want to study chain broken cases separately from chain corrected once consider submitting the physical problem (draw physical samples over all qubits, rather than logical samples), and separate out the broken from the unbroken manually.
    3) In most problems the objective e.g. mean-energy/ground-state-probability is a smooth function of the chain strength. If there is a strong dependence, one can do a bisection search over chain strength, to minimize the objective.
    4) As for 1-3 one has to first decide the meaning of response quality. As mentioned above, one is probably interested in finding say the minimal energy state for fixed time resources, or minimizing the time to achieve the optima. The D-Wave device has a number of parameters for improving these and the improvements possible depend a lot on the problem, in context of embedding problems: anneal duration and chain strength are perhaps the two ones you should begin with, trying to simplify your problem (e.g. by zeroing out some irrelevant couplers, simplifying the embedding) can also be very effective.

    1
    Comment actions Permalink

Please sign in to leave a comment.

Didn't find what you were looking for?

New post