Chapter 11· Input Selection for an Aigal Bloom Model
219
at an above average frequency. This mechanism ensures that a uniform data
distribution develops in the Kohonen layer. In adjusting the dista~ce, a bias, B p is
added to the distance and forms the new adjusted distance, D t The bias is
calculated using
(11.3)
where ' Y is a learning coefficient; F j is the frequency at which the PE j has
historically won; and M is the number o~ PEs in the Kohonen layer. ance B j and
D j are computed, the adjusted distance, D j can be calculated using
(11.4)
To ensure biological plausibility, lateral interaction with neighbouring PEs is
enforced by applying arbitrary network structures called neighbourhood sets, Ne'
Throughout the process, all PEs within the winner's neighbourhood set will have
their weights updated, whilst PEs outside of this set are left intact. The width or
radius of Ne can be time variable. The updating process to implement this
procedure is given by
(11.5)
where • is a scalar valued adaptation gain O<·(t)· ... and Ne is the neighbourhood
set. After the weights have been updated, the next input is presented to the
network and the process continues until convergence has been reached. After
successively presenting different inputs to the SaM, the net effect is that the
weights reflect the topological relationship that exists within the input data (Islam
and Kothari 2000).
Implementation ofthe SOM
The SaM has been used in ecological modelling applications to order data by
similarity (e.g. Chon et al. 1996; Foody 1999). In this paper, the SaM is used to
cluster the input variables into groups of similar inputs. By then sampling one
input from each cluster, it is possible to remove highly correlated, redundant
variables from the original data set. The SaM is implemented using the NCS
NeuFrame software. To cluster the data, the input variables are presented to the
network as the SaM's inputs. The software default parameters are used for the
learning rate, neighbourhood size and number of epochs. The output of the SaM
is obtained using aDynamie Patterns grid, wh ich shows adynamie representation
of the nodes that are winning each pattern. Each individual cell in the grid
represents anode in the Kohonen layer. There is no theoretical principle for
determining the optimum size of the Kohonen layer (Cai et al. 1994), hence, the
Kohonen layer was kept large enough to ensure that the maximum number of
219
at an above average frequency. This mechanism ensures that a uniform data
distribution develops in the Kohonen layer. In adjusting the dista~ce, a bias, B p is
added to the distance and forms the new adjusted distance, D t The bias is
calculated using
(11.3)
where ' Y is a learning coefficient; F j is the frequency at which the PE j has
historically won; and M is the number o~ PEs in the Kohonen layer. ance B j and
D j are computed, the adjusted distance, D j can be calculated using
(11.4)
To ensure biological plausibility, lateral interaction with neighbouring PEs is
enforced by applying arbitrary network structures called neighbourhood sets, Ne'
Throughout the process, all PEs within the winner's neighbourhood set will have
their weights updated, whilst PEs outside of this set are left intact. The width or
radius of Ne can be time variable. The updating process to implement this
procedure is given by
(11.5)
where • is a scalar valued adaptation gain O<·(t)· ... and Ne is the neighbourhood
set. After the weights have been updated, the next input is presented to the
network and the process continues until convergence has been reached. After
successively presenting different inputs to the SaM, the net effect is that the
weights reflect the topological relationship that exists within the input data (Islam
and Kothari 2000).
Implementation ofthe SOM
The SaM has been used in ecological modelling applications to order data by
similarity (e.g. Chon et al. 1996; Foody 1999). In this paper, the SaM is used to
cluster the input variables into groups of similar inputs. By then sampling one
input from each cluster, it is possible to remove highly correlated, redundant
variables from the original data set. The SaM is implemented using the NCS
NeuFrame software. To cluster the data, the input variables are presented to the
network as the SaM's inputs. The software default parameters are used for the
learning rate, neighbourhood size and number of epochs. The output of the SaM
is obtained using aDynamie Patterns grid, wh ich shows adynamie representation
of the nodes that are winning each pattern. Each individual cell in the grid
represents anode in the Kohonen layer. There is no theoretical principle for
determining the optimum size of the Kohonen layer (Cai et al. 1994), hence, the
Kohonen layer was kept large enough to ensure that the maximum number of