Commit 241e0a425d39430d9950532d37d68df9bfe998d1

  • avatar
  • Matthieu Weber <mweber @m…t.jyu.fi> (Committer)
  • Wed Jul 21 16:18:20 EEST 2010
  • avatar
  • Matthieu Weber <mweber @m…t.jyu.fi> (Author)
  • Wed Jul 21 16:18:20 EEST 2010
Modifications according to reviewers' comments
histograms.tex
(25 / 11)
  
102102shows that the table requires no more space than the traditional one, and
103103because it uses a graphical representation rather than a textual one, it is
104104instantly readable.
105%
106%The remainder of this article is organized as follows:
107%Section~\ref{s:shortcomings} presents the shortcomings of current practices,
108%Section~\ref{s:sparklines} describes the concept of sparkline histogram,
109%Section~\ref{s:stats} discusses the difficulties of comparing optimization
110%methods, Section~\ref{s:comparison} compares stacked focused histograms to
111%other ways of presenting experimental data and Section~\ref{s:conc} concludes
112%this paper.
105113%}}}1
106114
107115\section{\uppercase{Shortcomings of current practices}}\label{s:shortcomings}%{{{1
141141the time to read the numbers carefully and compare them.
142142%}}}1
143143
144\section{\uppercase{The Sparkline histograms}} %{{{1
144\section{\uppercase{The Sparkline histograms}}\label{s:sparklines} %{{{1
145145
146146% Since the goal of an author is to describe a set of values and that we ruled
147147% out the use of any estimator such as the average as not significant on samples
333333a full-size example of stacked, focused sparklines and is compared against the
334334traditional table of average values (in Table~ \ref{tab:stdavg}; see
335335Section~\ref{s:comparison} for a detailed description of those tables).
336
337Reading a sparkline histogram is a three step procedure. First the reader
338needs to verify the range, which gives him a rough estimate of the scale of
336Reading a sparkline histogram is a three step procedure:
337\begin{enumerate}
338\item Verify the range, which gives a rough estimate of the scale of
339339the values, serving the same function as the average value, or the y-axis of a
340340convergence graph. If the range is outside of what the reader considers
341interesting, the rest of the graph can be discarded. The reader should focus
342next on the dump bin. Those algorithms that are entirely, or almost entirely
343dumped can be discarded as uninteresting. The remaining part is the one the
344author of the graphic has deemed interesting.
341interesting, the rest of the graph can be discarded.
342\item Focus on the dump bin. Those algorithms that are entirely, or almost
343entirely dumped can be discarded as uninteresting.
344\item Draw conclusions from the remaining part, which is the one the author of
345the graphic has deemed interesting.
346\end{enumerate}
345347
346348To create a stack of histograms as presented above, the following procedure
347349has been applied. It must be noted that this procedure assumes that
393393the same mean statistic, but one has larger deviation. In the above perspective of
394394comparing means, we would have to conclude that they are equal, in practise however,
395395the difference can be crucial as we can run the highly varying
396algorithm several times to ensure significantly better results. Matters are made
397even more complicated by skewed distributions, where we would need to observe even
396algorithm several times to ensure significantly better results. Skewed
397distributions complicate the matter even further since
398%Matters are made
399%even more complicated by skewed distributions, where
400we would need to observe
401%even
398402more statistical parameters in addition to variance to decide which algorithm would
399403be better at the case at hand.
400404
527527
528528%}}}1
529529
530\section{\uppercase{Conclusions}} % half a page
530\section{\uppercase{Conclusions}}\label{s:conc}%{{{1 % half a page
531531
532532In this text we have presented a novel visualization for comparing evolutionary optimization methods. We claim that this visualisation
533533can convey more information than average/standard deviation tables and statistical test tables while retaining nearly the same usage of