What is the difference between different embedding composites?

• 

  Sukhraan S  (Report)

  December 22, 2020 01:16

  Hi Aniruddha,

  Let's address your second question first.

  Structured samplers, such as the DWaveSampler, are restricted to sampling on a specific graph. Unstructured samplers, such as the ExactSolver, work with an arbitrary graph.

  Embedding Composites map a problem defined on one graph onto a different graph. One way we use them is for embedding problems onto our quantum processing units (QPUs). The different embedding composites we provide in Ocean are designed with different use cases in mind: 

  • AutoEmbeddingComposite attempts to run the problem directly on the QPU and only embeds if a BinaryQuadraticModelStructureError is thrown. It can be used with problems that appear to already have the same structure as the QPU architecture.
  • LazyFixedEmbeddingComposite calculates an embedding on the first call and reuses it on subsequent calls even if linear and quadratic weights are changed.
  • EmbeddingComposite calculates a new embedding automatically for every call to the sample function. It is the most commonly used composite.
  • FixedEmbeddingComposite samples from a user provided embedding. This is useful for large or complex problems that take a lot of time to embed or require a specifically tuned embedding.
  • TilingComposite tiles small problems across the QPU which allows for sampling multiple instances of a problem in parallel. 
  • VirtualGraphComposite optimizes the embedding using the advanced features of a given QPU. Note that this uses QPU time.
    
    

  Additional resources: 

  Intro to Minor embedding: https://docs.dwavesys.com/docs/latest/c_gs_7.html

  Embedding Composites: https://docs.ocean.dwavesys.com/en/stable/docs_system/reference/composites.html#embedding

  Reusing your embedding: https://support.dwavesys.com/hc/en-us/community/posts/360016737274

  0

  Comment actions Permalink

Please sign in to leave a comment.