A method called simulated breeding is applied to the problem of learning criteria for similarity-based image retrieval. Usually traditional genetic algorithms without any modifications are used in the simulated breeding methods; however we propose a new model ``pipeline-type genetic algorithms'' that is more suitable for the simulated breeding methods, and also discuss genetic operators that are special to this model. We then show that this model can eliminate both time constraints and the restriction of population size, which are the disadvantage of the simulated breeding methods, and also demonstrate this model's powerful search capability and effective parallel processing. Finally we describe the application of this model to similarity-based image retrieval as a good man-machine interface.

      @InProceedings{ kt:hip96,
      author = {Asanobu KITAMOTO and Mikio TAKAGI},
      title = {Learning Criteria for Similarity-Based Image Retrieval Using Simulated Breeding by Pipeline-Type Genetic Algorithms},
      booktitle = {Technical Report of Institute of Electronics, Information, and Communication Engineers (IEICE)},
      volume = {HIP96-4},
      pages = {17-22},
      year = 1996,
      month = 6,
      doi = {10.11485/tvtr.20.38_17},
      note = { (in Japanese)},
      }