Thesis or Dissertation Learning Classifier System Design based on Learning Strategy

中田, 雅也  ,  ナカタ, マサヤ  ,  Nakata, Masaya

pp.1 - 209 , 2015-09-30 , The University of Electro-Communications
On Learning Classifier Systems dubbed LCSs a leaning strategy which defines how LCSs cover a state-action space in a problem can be one of the most fundamental options in designing LCSs. There lacks an intensive study of the learning strategy to understand whether and how the learning strategy affects the performance of LCSs. This lack has resulted in the current design methodology of LCS which does not carefully consider the types of learning strategy. The thesis clarifies a need of a design methodology of LCS based on the learning strategy. That is, the thesis shows the learning strategy can be an option that determines the potential performance of LCSs and then claims that LCSs should be designed on the basis of the learning strategy in order to improve the performance of LCSs. First, the thesis empirically claims that the current design methodology of LCS, without the consideration of learning strategy, can be limited to design a proper LCS to solve a problem. This supports the need of design methodology based on the learning strategy. Next, the thesis presents an example of how LCS can be designed on the basis of the learning strategy. The thesis empirically show an adequate learning strategy improving the performance of LCS can be decided depending on a type of problem difficulties such as missing attributes. Then, the thesis draws an inclusive guideline that explains which learning strategy should be used to address which types of problem difficulties. Finally, the thesis further shows, on an application of LCS for a human daily activity recognition problem, the adequate learning strategy according to the guideline effectively improves the performance of the application. The thesis concludes that the learning strategy is the option of the LCS design which determines the potential performance of LCSs. Thus, before designing any type of LCSs including their applications, the learning strategy should be adequately selected at first, because their performance degrades when they employ an inadequate learning strategy to a problem they want to solve. In other words, LCSs should be designed on the basis of the adequate learning strategy.

Number of accesses :  

Other information