Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 198

7.4.3 Darwin Neural Network
In this section, we explain a structure adaptive learning algorithm for neural
networks based on the theory of evolution [4]. Our proposed method imitates the
process that living things adapt their structures according to the environment by
evolution and learning. In this method, if the teaching data change during learning
under dynamic environments, the learning does not restart from the initial state.
This method is useful for adaptive learning, which can take into account
inheritance of the network structure, the connection weight vectors, and the
learning parameters.
From the theory of evolution, Sasaki and Tokoro compared two types of
hereditary mechanisms: Lamarckian and Darwinian [16], [17]. There are two
phases in each mechanism: the evolution phase over generations and the learning
phase in the individual's lifetime. In [16] and [17], they regarded neural network's
structure as the individual in population. BP learning is employed as the learning
method and GA search is employed as the evolution method; the two algorithms
work together to adapt their parameters under their environments. The Lamarckian
hereditary mechanism in Fig. 7.15 inherits the trained connection weights by BP
learning to the next generation. The Darwinian hereditary mechanism in Fig. 7.16
inherits only chromosomes from their parents. They reported that the performance
of Darwinian is better than that of Lamarckian under their assumptions.
The evolution in the real world includes the change of length of chromosome,
but their approach does not account for it. Also, some parameters in BP learning
and GA search were determined by the designer. To improve this, we describe our
proposed method in the next section.
Fig.7.15. Hereditary mechanism of Lamarckian type.