demogng.tex

关于自组织神经网络的一种新结构程序,并包含了其它几种神经网络的程序比较

%\caption{Overview of the available probability distributions} 
%\end{center} 
%\end{figure}


\subsection{Model Specific Options}

This region shows the model specific parameters. Each time a new model
is selected, the necessary parameters are displayed. For a complete
description of the models and their parameters look at the technical report
\htmladdnormallink{\textit{Some Competitive Learning Methods}}{http://www.neuroinformatik.ruhr-uni-bochum.de/ini/VDM/research/gsn/JavaPaper}
by Bernd Fritzke which is available in
\htmladdnormallinkfoot{HTML}{http://www.neuroinformatik.ruhr-uni-bochum.de/ini/VDM/research/gsn/JavaPaper}-
and in 
\htmladdnormallinkfoot{Postscript-format}{ftp://ftp.neuroinformatik.ruhr-uni-bochum.de/pub/software/NN/DemoGNG/sclm.ps.gz}.

%
% Change font in description from bold to typewriter
%
\renewcommand{\descriptionlabel}[1]{\hspace{\labelsep}\fbox{\textbf{#1}}}

\subsubsection{\simImDGNG{LBG}{LBG_7}, \simDGNG{LBG-U}{LBG-U_7}}
\begin{description}
\item[Number of Signals] The number of input signals (only discrete
  distributions).
\item[LBG-U] Switches from LBG to LBG-U and back.
\end{description}

\subsubsection{\simImDGNG{Hard Competitive Learning}{HCL_5}}
\begin{description}
\item[Variable] Switches from a constant to a variable learning rate.
\item[epsilon] This value ($\epsilon$) determines the extent to which the
  winner is adapted towards the input signal (constant learning rate).
\item[epsilon\_i] epsilon initial ($\epsilon_i$).
\item[epsilon\_f] epsilon final ($\epsilon_f$).
\item[t\_max] The simulation ends, if the number of input signals
  exceeds this value ($t_{max}$).
\end{description}

The variable learning rate is determined according to
\begin{displaymath}
  \qquad\epsilon(t) = \epsilon_i(\epsilon_f/\epsilon_i)^{t/t_{\rm max}}.
\end{displaymath}


\subsubsection{\simImDGNG{Neural Gas}{NG_3}}
\begin{description}
\item[lambda\_i] lambda initial ($\lambda_i$).
\item[lambda\_f] lambda final ($\lambda_f$).
\item[epsilon\_i] epsilon initial ($\epsilon_i$).
\item[epsilon\_f] epsilon final ($\epsilon_f$).
\item[t\_max] The simulation ends, if the number of input signals
  exceeds this value ($t_{max}$).
\end{description}
The reference vectors are adjusted according to
\begin{displaymath}
  \Delta \bw_i = \epsilon(t) \cdot h_\lambda(k_i(\bxi,\A)) \cdot (\bxi - \bw_i)
\end{displaymath} 
with the following time-dependencies:
\begin{displaymath}
  \lambda(t) = \lambda_i (\lambda_f/\lambda_i)^{t/t_{\rm max}}
\end{displaymath}
\begin{displaymath}
  \qquad\epsilon(t) = \epsilon_i(\epsilon_f/\epsilon_i)^{t/t_{\rm max}}
\end{displaymath}
\begin{displaymath}
  h_\lambda(k) = \exp(-k/\lambda(t)).
\end{displaymath}

\subsubsection{\simImDGNG{Competitive Hebbian Learning}{CHL_0}}
This implementation requires no  model specific parameters. In general
one would have a maximum number of time steps ($t_{max}$).

\subsubsection{\simImDGNG{Neural Gas with Competitive Hebbian Learning}{NGwCHL_2}}
\begin{description}
\item[lambda\_i] lambda initial ($\lambda_i$).
\item[lambda\_f] lambda final ($\lambda_f$).
\item[epsilon\_i] epsilon initial ($\epsilon_i$).
\item[epsilon\_f] epsilon final ($\epsilon_f$).
\item[t\_max] The simulation ends, if the number of input signals
  exceed this value ($t_{max}$).
\item[edge\_i] Initial value for time-dependend edge aging ($T_i$).
\item[edge\_f] Final value for time-dependend edge aging ($T_f$).
\end{description}
Edges are removed with an age larger than the maximal age $T(t)$ whereby
\begin{displaymath}
T(t) = T_i(T_f/T_i) ^ {t/t_{\rm max}}.
\end{displaymath}
The reference vectors are adjusted according to
\begin{displaymath}
  \Delta \bw_i = \epsilon(t) \cdot h_\lambda(k_i(\bxi,\A)) \cdot (\bxi - \bw_i)
\end{displaymath} 
with the following time-dependencies:
\begin{displaymath}
  \lambda(t) = \lambda_i (\lambda_f/\lambda_i)^{t/t_{\rm max}}
\end{displaymath}
\begin{displaymath}
  \qquad\epsilon(t) = \epsilon_i(\epsilon_f/\epsilon_i)^{t/t_{\rm max}}
\end{displaymath}
\begin{displaymath}
  h_\lambda(k) = \exp(-k/\lambda(t)).
\end{displaymath}

\subsubsection{\simImDGNG{Growing Neural Gas}{GNG_3}, \simImDGNG{Growing Neural
    Gas with Utility}{GNG-U_10}}
\begin{description}
\item[No new Nodes] No new nodes will be inserted.
\item[Utility] Switches from GNG to GNG-U and back. The value ($k$) determines
  the deletion of the unit with the smallest utility $U_i$ ($i := arg\ 
  min_cU_c$), if the utility value falls below a certain fraction of the error
  variable $E_q$: $k < E_q/U_i$.
\item[Lambda] If the number of input signals generated so far is an
  integer multiple of this value ($\lambda$), insert a new node.
\item[max.~Edge Age] Remove edges with an age larger than this value
  ($a_{max}$). If this results in nodes having no emanating edges,
  remove them as well.
\item[Epsilon winner] Move the winner node towards the input signal by this
  fraction of the total distance ($\epsilon_b$).
\item[Epsilon neighbor] Move the neighbors of the winner node towards the
  input signal by this fraction of the total distance ($\epsilon_n$).
\item[alpha] Decrease the error variables of the nodes neighboring to the newly
  inserted node by a fraction of this size ($\alpha$).
\item[beta] Decrease the error variables of all nodes by a fraction
  of this size ($\beta$).
\end{description}

\subsubsection{\simImDGNG{Self-Organizing Map}{SOM_8}}
\begin{description}
\item[Grid size] The form of the grid and the number of nodes are determined.
\item[epsilon\_i] epsilon initial ($\epsilon_i$).
\item[epsilon\_f] epsilon final ($\epsilon_f$).
\item[sigma\_i] sigma initial ($\sigma_i$).
\item[sigma\_f] sigma final ($\sigma_f$).
\item[t\_max] The simulation ends, if the number of input signals
  exceed this value ($t_{max}$).
\end{description}
Determine $\epsilon(t)$ and $\sigma(t)$ according to
\begin{displaymath}
\epsilon(t) = \epsilon_i(\epsilon_f/\epsilon_i)^{t/t_{\rm max}},
\end{displaymath}
\begin{displaymath}
\sigma(t) = \sigma_i(\sigma_f/\sigma_i)^{t/t_{\rm max}}.
\end{displaymath}

\subsubsection{\simImDGNG{Growing Grid}{GG_8}}
\begin{description}
\item[No new Nodes]  No new nodes will be inserted.
\item[lambda\_g] This parameter ($\lambda_g$) indicates how many
  adaptation steps on average are done per node before new nodes are
  inserted (\textit{growth phase}).
\item[lambda\_f]  This parameter ($\lambda_f$) indicates how many
  adaptation steps on average are done per node before new nodes are
  inserted (\textit{fine-tuning phase}).
\item[epsilon\_i] epsilon initial ($\epsilon_i$).
\item[epsilon\_f] epsilon final ($\epsilon_f$).
\item[sigma] This parameter ($\sigma$) determines the width of the
  bell-shaped neighborhood interaction function.
\end{description}
In the fine-tuning phase the time-dependend learning rate $\epsilon(t)$
is determined according to
\begin{displaymath}
  \qquad\epsilon(t) = \epsilon_i(\epsilon_f/\epsilon_i)^{t/t_{\rm max}}.
\end{displaymath}

\section{Wishlist}
This is a list of projects for DemoGNG. Bug reports and requests which can
not be fixed or honored right away will be added to this list. If you have
some time for Java hacking, you are encouraged to try to provide a
solution to one of the following problems. It might be a good idea to
send a mail first, though.

\begin{itemize}
\item Change looping in the \texttt{run}-procedure.
\item New Method $k$-means~\cite{MacQueen67}.
\item Redesign of the user interface (JDK 1.2).
\item Supervised Learning
\item Tune the main \texttt{Makefile}.
\end{itemize}

Please send any comments or suggestions you might have, along with
any bugs that you may encounter to:\newline \newline
Hartmut S. Loos\newline
\htmladdnormallink{[e-mail]}{mailto:[e-mail]}

\newpage
%\hfill

\addcontentsline{toc}{section}{References}
%\bibliographystyle{alpha}
\bibliographystyle{apalike}
\bibliography{DemoGNG}

\newpage

\section{Change log}

\begin{verbatim}
19.10.1998: Version 1.5 released.
\end{verbatim}

These are the most important changes from v1.0 to v1.5:

\begin{verbatim}
19.10.1998: Updated the manual.
17.07.1998: LBG: Marked nodes which haven't moved since the last
            update. 
16.07.1998: GNG-U: Fixed a small bug regarding 'beta' for utility. 
15.07.1998: Non-stationary probability distributions. 
23.10.1997: Fixed: nodes without neighbors are not deleted in GNG-U
            (but in GNG). 
13.10.1997: Trifles in GNG-U. 
29.09.1997: Fixed a small bug in Method LBG. 
28.09.1997: New Method LBG-U (Fritzke 1997). 
27.09.1997: New Method GNG-U (Fritzke 1997).

10.03.1997: Version 1.3 released.

09.03.1997: Proofred the manual.
08.03.1997: Changed the order of the models (model menu
            and manual).
07.03.1997: GNG: added two variables (alpha, beta).
06.03.1997: Tuned the code (20% faster).
05.03.1997: Cleaned up the code/deleted DEBUG statements.
04.03.1997: Updated the manual.
03.03.1997: Added a new button 'Random Init'.
02.03.1997: Made some improvements concerning LBG.

28.02.1997: Version 1.2 released.
26.02.1997: Version 1.1 skipped for PR reasons :-).

24.02.1997: Added description of the model specific options to
            the manual.
24.02.1997: Improved the variable learning rate for Hard
            Competitive Learning.
21.02.1997: Changed eta to epsilon.
20.02.1997: First draft of the manual for v1.1.
18.02.1997: Updated the manual.
17.02.1997: Made some screen dumps for the manual.
30.10.1996: Some improvements and a bugfix concerning the error
            values. 
28.10.1996: New method Self-organizing map (SOM).
23.10.1996: Redesigned the user interface. I have made minor
            changes only. The next major version 2.0 will have
            a complete new interface. 
22.10.1996: Added a close button to the error graph.
21.10.1996: In the method Growing Grid a percentage counter appears
            in the fine-tuning phase. At 100% the calculation stops.
17.10.1996: Added a new button 'White'. It switches the background
            of the drawing area to white. This is useful for making
            hardcopies of the screen.
17.10.1996: New distribution UNIT.
16.10.1996: Fixed a bug: Now the Voronoi diagram/Delaunay
            triangulation is also available for Growing Grid.
15.10.1996: New method Growing Grid.
14.10.1996: Added a new class GridNodeGNG to gain a grid covering
            the nodes. 
09.10.1996: Added a new menu for speed. Now it is possible to switch
            to an individual speed depending on the machine and/or
            browser. (On some machines there was no interaction
            possible with the default value. On the other hand, why
            longer waiting than necessary?) 
04.10.1996: Added an error graph. (The source of this class was
            written by Christian Kurzke and Ningsui Chen. I have
            only made minor changes to this class. Many thanks to
            Christian Kurzke and Ningsui Chen for this nice
            graph class.)
30.09.1996: Added a new frame in a separate window for the error
            graph (new class GraphGNG).
30.09.1996: New checkboxes to turn the nodes/edges on and off.
26.09.1996: Fixed again all known bugs.
25.09.1996: Finished the discrete distribution (mainly for LBG).
24.09.1996: Default no sound.
20.09.1996: Fixed all known bugs (Can you find some unknown?).
19.09.1996: Prepared for new distribution (a discrete one with 500
            fixed signals).
10.09.1996: New method LBG (some minor bugs are known, but not
            concerning the method).
03.09.1996: Renamed class PanelGNG to a more convenient name
            (ComputeGNG).
02.09.1996: Split the source file DemoGNG.java. Now each class has
            its own file.
01.09.1996: Inserted more comments.
30.08.1996: Cleaned up the code to compute the Voronoi
            diagram/Delaunay triangulation.
07.08.1996: Added Delaunay triangulation (press checkbutton
            Delaunay)!
06.08.1996: Now you can display Voronoi diagrams for each method
            (press checkbutton Voronoi)!

21.06.1996: Version 1.0 released.
\end{verbatim}

\end{document}