<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <p>It might make sense to enable resignation threshold even on
      stupid level. As such the first thing the network should learn
      would be not to resign to early (even before not passing)<br>
    </p>
    <br>
    <div class="moz-cite-prefix">Le 02/12/2017 à 18:17, Brian Sheppard
      via Computer-go a écrit :<br>
    </div>
    <blockquote type="cite"
      cite="mid:117101d36b91$68acee70$3a06cb50$@aol.com">
      <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
      <meta name="Generator" content="Microsoft Word 15 (filtered
        medium)">
      <style><!--
/* Font Definitions */
@font-face
        {font-family:"Cambria Math";
        panose-1:2 4 5 3 5 4 6 3 2 4;}
@font-face
        {font-family:Calibri;
        panose-1:2 15 5 2 2 2 4 3 2 4;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
        {margin:0in;
        margin-bottom:.0001pt;
        font-size:12.0pt;
        font-family:"Times New Roman",serif;}
a:link, span.MsoHyperlink
        {mso-style-priority:99;
        color:blue;
        text-decoration:underline;}
a:visited, span.MsoHyperlinkFollowed
        {mso-style-priority:99;
        color:purple;
        text-decoration:underline;}
span.EmailStyle17
        {mso-style-type:personal;
        font-family:"Calibri",sans-serif;
        color:#1F497D;}
span.EmailStyle18
        {mso-style-type:personal-reply;
        font-family:"Calibri",sans-serif;
        color:#1F497D;}
.MsoChpDefault
        {mso-style-type:export-only;
        font-size:10.0pt;}
@page WordSection1
        {size:8.5in 11.0in;
        margin:1.0in 1.0in 1.0in 1.0in;}
div.WordSection1
        {page:WordSection1;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
      <div class="WordSection1">
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">I
            have some hard data now. My network’s initial training
            reached the same performance in half the iterations. That
            is, the steepness of skill gain in the first day of training
            was twice as great when I avoided training on fill-ins.<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">The
            has all the usual caveats: only one run before/after, YMMV,
            etc.<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <div>
          <div style="border:none;border-top:solid #E1E1E1
            1.0pt;padding:3.0pt 0in 0in 0in">
            <p class="MsoNormal"><b><span
                  style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif">
                Brian Sheppard [<a class="moz-txt-link-freetext" href="mailto:sheppardco@aol.com">mailto:sheppardco@aol.com</a>] <br>
                <b>Sent:</b> Friday, December 1, 2017 5:39 PM<br>
                <b>To:</b> 'computer-go'
                <a class="moz-txt-link-rfc2396E" href="mailto:computer-go@computer-go.org"><computer-go@computer-go.org></a><br>
                <b>Subject:</b> RE: [Computer-go] Significance of
                resignation in AGZ<o:p></o:p></span></p>
          </div>
        </div>
        <p class="MsoNormal"><o:p> </o:p></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">I
            didn’t measure precisely because as soon as I saw the
            training artifacts I changed the code. And I am not doing an
            AGZ-style experiment, so there are differences for sure. So
            I will give you a swag…<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Speed
            difference is maybe 20%-ish for 9x9 games.<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">A
            frequentist approach will overstate the frequency of fill-in
            plays by a pretty large factor, because fill-in plays are
            guaranteed to occur in every game but are not best in the
            competitive part of the game. This will affect the speed of
            learning in the early going.<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">The
            network will use some fraction (almost certainly <= 20%)
            of its capacity to improve accuracy on positions that will
            not contribute to its ultimate strength. This applies to
            both ordering and evaluation aspects.<o:p></o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
        <p class="MsoNormal"><b><span
              style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span
style="font-size:11.0pt;font-family:"Calibri",sans-serif">
            Andy [<a class="moz-txt-link-freetext" href="mailto:andy.olsen.tx@gmail.com">mailto:andy.olsen.tx@gmail.com</a>] <br>
            <b>Sent:</b> Friday, December 1, 2017 4:55 PM<br>
            <b>To:</b> Brian Sheppard <a class="moz-txt-link-rfc2396E" href="mailto:sheppardco@aol.com"><sheppardco@aol.com></a>;
            computer-go <a class="moz-txt-link-rfc2396E" href="mailto:computer-go@computer-go.org"><computer-go@computer-go.org></a><br>
            <b>Subject:</b> Re: [Computer-go] Significance of
            resignation in AGZ<o:p></o:p></span></p>
        <p class="MsoNormal"><o:p> </o:p></p>
        <div>
          <p class="MsoNormal">Brian, do you have any experiments
            showing what kind of impact it has? It sounds like you have
            tried both with and without your ad hoc first pass approach?<o:p></o:p></p>
          <div>
            <p class="MsoNormal"><o:p> </o:p></p>
          </div>
          <div>
            <p class="MsoNormal"><o:p> </o:p></p>
          </div>
          <div>
            <div>
              <p class="MsoNormal"><o:p> </o:p></p>
            </div>
          </div>
        </div>
        <div>
          <p class="MsoNormal"><o:p> </o:p></p>
          <div>
            <p class="MsoNormal">2017-12-01 15:29 GMT-06:00 Brian
              Sheppard via Computer-go <<a
                href="mailto:computer-go@computer-go.org"
                target="_blank" moz-do-not-send="true">computer-go@computer-go.org</a>>:<o:p></o:p></p>
            <blockquote style="border:none;border-left:solid #CCCCCC
              1.0pt;padding:0in 0in 0in
6.0pt;margin-left:4.8pt;margin-top:5.0pt;margin-right:0in;margin-bottom:5.0pt">
              <p class="MsoNormal">I have concluded that AGZ's policy of
                resigning "lost" games early is somewhat significant.
                Not as significant as using residual networks, for sure,
                but you wouldn't want to go without these advantages.<br>
                <br>
                The benefit cited in the paper is speed. Certainly a
                factor. I see two other advantages.<br>
                <br>
                First is that training does not include the "fill in"
                portion of the game, where every move is low value. I
                see a specific effect on the move ordering system, since
                it is based on frequency. By eliminating training on
                fill-ins, the prioritization function will not be biased
                toward moves that are not relevant to strong play. (That
                is, there are a lot of fill-in moves, which are usually
                not best in the interesting portion of the game, but
                occur a lot if the game is played out to the end, and
                therefore the move prioritization system would predict
                them more often.) My ad hoc alternative is to not train
                on positions after the first pass in a game. (Note that
                this does not qualify as "zero knowledge", but that is
                OK with me since I am not trying to reproduce AGZ.)<br>
                <br>
                Second is the positional evaluation is not training on
                situations where everything is decided, so less of the
                NN capacity is devoted to situations in which nothing
                can be gained.<br>
                <br>
                As always, YMMV.<br>
                <br>
                Best,<br>
                Brian<br>
                <br>
                <br>
                _______________________________________________<br>
                Computer-go mailing list<br>
                <a href="mailto:Computer-go@computer-go.org"
                  moz-do-not-send="true">Computer-go@computer-go.org</a><br>
                <a
                  href="http://computer-go.org/mailman/listinfo/computer-go"
                  target="_blank" moz-do-not-send="true">http://computer-go.org/mailman/listinfo/computer-go</a><o:p></o:p></p>
            </blockquote>
          </div>
          <p class="MsoNormal"><o:p> </o:p></p>
        </div>
      </div>
      <br>
      <fieldset class="mimeAttachmentHeader"></fieldset>
      <br>
      <pre wrap="">_______________________________________________
Computer-go mailing list
<a class="moz-txt-link-abbreviated" href="mailto:Computer-go@computer-go.org">Computer-go@computer-go.org</a>
<a class="moz-txt-link-freetext" href="http://computer-go.org/mailman/listinfo/computer-go">http://computer-go.org/mailman/listinfo/computer-go</a></pre>
    </blockquote>
    <br>
  </body>
</html>