HELP CLEDITMENU                             Julian Clinton  Feb 1990


This file describes the options available on the Edit Competitive
Learning Network Menu.


         CONTENTS - (Use <ENTER> g to access required sections)

 -- 'learning rate for winning units'
 -- 'learning rate for losing units'
 -- 'sensitivity equalisation for winning units'
 -- 'sensitivity equalisation for losing units'


-- 'learning rate for winning units' ----------------------------------

When the competition between units in a group has been decided, the
strengths of connections between the winning units and all winning units
in the previous layer are increased by a proportion defined this
parameter.


-- 'learning rate for losing units' -----------------------------------

While the winning unit in a group strengthens its connections to winning
units in the previous layer, the losing units in a group can strengthen
their connections to losing units in the previous layer to prevent the
situation where one unit dominates in each group. This is called leaky
learning and can be very expensive computationally. You can set this
variable to FALSE to switch off leaky learning.


-- 'sensitivity equalisation for winning units' -----------------------

To prevent a winning unit from dominating a group, the sensitivity of
the winning unit can be reduced. Each unit has a "resting activation"
called a "bias" which is reduced by a factor defined by this option
whenever the unit wins. This can be set to FALSE.


-- 'sensitivity equalisation for losing units' ------------------------

As winning units can become less sensitive, so losing units can become
more sensitive to increase their chance of winning next time. The factor
by which the bias of losing units is increased can be changed by
selecting this option (which can also be set to FALSE).


--- Copyright Integral Solutions Ltd. 1990. All rights reserved. ---
