|
|
|
|
— |
speed_of_ai_transition:range_of_human_performance:time_for_ai_to_cross_the_human_performance_range_in_go [2022/09/21 07:37] (current)
|
| | ====== Time for AI to cross the human performance range in Go ====== |
| | |
| | // Published 15 October, 2020; last updated 08 March, 2021 // |
| | |
| | |
| | |
| | <HTML> |
| | <p>Progress in computer Go performance took:</p> |
| | </HTML> |
| | |
| | |
| | <HTML> |
| | <ul> |
| | <li><div class="li">0-19 years to go from the first attempt to playing at human beginner level (<1987)</div></li> |
| | <li><div class="li">>30 years to go from human beginner level to superhuman level (<1987-2017)</div></li> |
| | <li><div class="li">3 years to go from superhuman level to the the current highest performance (2017-2020)</div></li> |
| | </ul> |
| | </HTML> |
| | |
| | |
| | |
| | ===== Details ===== |
| | |
| | |
| | ==== Human performance milestones ==== |
| | |
| | |
| | <HTML> |
| | <p>Human go ratings range from 30 kyu (beginner), through 7 dan to at least 9 professional dan.<span class="easy-footnote-margin-adjust" id="easy-footnote-1-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-1-2680" title='“Go Ranks and Ratings.” In <em>Wikipedia</em>, June 20, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&amp;oldid=963489455">https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&amp;oldid=963489455</a>.'><sup>1</sup></a></span> These ratings go downwards through kyu levels, then upward through dan levels, then upward through professional dan levels. The top ratings seem to be <a href="http://en.wikipedia.org/wiki/Go_ranks_and_ratings#Elo-like_rating_systems_as_used_in_Go">closer together</a> than the lower ones, though there are apparently <a href="https://en.wikipedia.org/wiki/Go_ranks_and_ratings#Winning_probabilities">multiple systems</a> which vary)<span class="easy-footnote-margin-adjust" id="easy-footnote-2-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-2-2680" title='See table: <br><br>“Go Ranks and Ratings.” In <em>Wikipedia</em>, June 20, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&amp;oldid=963489455">https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&amp;oldid=963489455</a>.'><sup>2</sup></a></span></p> |
| | </HTML> |
| | |
| | |
| | ==== AI achievement of human milestones ==== |
| | |
| | |
| | === Earliest attempt === |
| | |
| | |
| | <HTML> |
| | <p>Wikipedia says the first Go program was written in 1968.<span class="easy-footnote-margin-adjust" id="easy-footnote-3-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-3-2680" title='&#8220;The first Go program was written by&nbsp;<a href="https://en.wikipedia.org/wiki/Albert_Lindsey_Zobrist">Albert Lindsey Zobrist</a>&nbsp;in 1968 as part of his thesis on&nbsp;<a href="https://en.wikipedia.org/wiki/Pattern_recognition">pattern recognition</a>.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-11">[11]</a></sup>&nbsp;It introduced an&nbsp;<a href="https://en.wikipedia.org/wiki/Influence_function_(statistics)">influence function</a>&nbsp;to estimate territory and&nbsp;<a href="https://en.wikipedia.org/wiki/Zobrist_hashing">Zobrist hashing</a>&nbsp;to detect&nbsp;<a href="https://en.wikipedia.org/wiki/Ko_rule">ko</a>.&#8221;'><sup>3</sup></a></span> We do not know how well it performed.</p> |
| | </HTML> |
| | |
| | |
| | === Beginner level === |
| | |
| | |
| | <HTML> |
| | <p>We have not investigated early Go performance in depth. Figure 1 includes informed guesses about early performance by David Fotland, author of successful Go program, <em>The Many Faces of Go</em>, and <em>Sensei’s Library</em>, a Go wiki.<span class="easy-footnote-margin-adjust" id="easy-footnote-4-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-4-2680" title="&#8220;Figure 25 shows estimates from two<br>sources: David Fotland—author of The Many Faces of Go, an Olympiad-winning Go program—and Sensei’s Library, a collaborative Go wiki. David Fotland warns that the data from before bots played on KGS is poor, as programs tended not to play in human tournaments and so failed to get ratings.&#8221;</p> <p>Grace, Katja. “Algorithmic Progress in Six Domains.” Berkeley, CA: Machine Intelligence Research Institute, 2013."><sup>4</sup></a></span> Fotland says that early data on AI Go performance is poor, since bots did not play in tournaments, so were not rated.</p> |
| | </HTML> |
| | |
| | |
| | <HTML> |
| | <figure class="wp-block-image size-large is-resized"> |
| | <img alt="" class="wp-image-372" height="417" sizes="(max-width: 576px) 100vw, 576px" src="https://aiimpacts.org/wp-content/uploads/2014/12/gobothistory-copy-1024x743.jpg" srcset="https://aiimpacts.org/wp-content/uploads/2014/12/gobothistory-copy-1024x743.jpg 1024w, https://aiimpacts.org/wp-content/uploads/2014/12/gobothistory-copy-300x218.jpg 300w, https://aiimpacts.org/wp-content/uploads/2014/12/gobothistory-copy.jpg 1358w" width="576"/> |
| | <figcaption> |
| | <strong>Figure 1</strong>: From <a href="http://intelligence.org/files/AlgorithmicProgress.pdf">Grace 2013</a>. |
| | </figcaption> |
| | </figure> |
| | </HTML> |
| | |
| | |
| | <HTML> |
| | <p>This suggests that by 1987 Go bots were performing better than human beginners. We do not have evidence to pin down the date of human beginner level AI better, but have also not investigated thoroughly (there appears to be more evidence).</p> |
| | </HTML> |
| | |
| | |
| | === Superhuman level === |
| | |
| | |
| | <HTML> |
| | <p>In May 2017 AlphaGo beat the top ranked Go player in the world.<span class="easy-footnote-margin-adjust" id="easy-footnote-5-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-5-2680" title='&#8220;In May 2017, AlphaGo beat&nbsp;<a href="https://en.wikipedia.org/wiki/Ke_Jie">Ke Jie</a>, who at the time was ranked top in the world,<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-27">[27]</a><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-28">[28]</a></sup>&nbsp;in a&nbsp;<a href="https://en.wikipedia.org/wiki/AlphaGo_versus_Ke_Jie">three-game match</a>&nbsp;during the&nbsp;<a href="https://en.wikipedia.org/wiki/Future_of_Go_Summit">Future of Go Summit</a>.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-wuzhensecond-29">[29]</a></sup>&#8220;<br><br>“Computer Go.” In <em>Wikipedia</em>, July 27, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Computer_Go&amp;oldid=969736537">https://en.wikipedia.org/w/index.php?title=Computer_Go&amp;oldid=969736537</a>'><sup>5</sup></a></span> This does not imply that AlphaGo was overall better, but a new version in October could beat the May version in 89 games out of 100<span class="easy-footnote-margin-adjust" id="easy-footnote-6-2680"></span><span class="easy-footnote"><a href="#easy-footnote-bottom-6-2680" title='&#8220;In October 2017, DeepMind revealed a new version of AlphaGo, trained only through self play, that had surpassed all previous versions, beating the Ke Jie version in 89 out of 100 games.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-alphagozero-30">[30]</a></sup>&#8220;</p> <p>“Computer Go.” In <em>Wikipedia</em>, July 27, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Computer_Go&amp;oldid=969736537">https://en.wikipedia.org/w/index.php?title=Computer_Go&amp;oldid=969736537</a>'><sup>6</sup></a></span>, suggesting that if in May it would have beaten Ke Jie in more than 11% of games, the new version would beat Ke Jie more than half the time, i.e. perform better than the best human player. Thus 2017 seems like a reasonable date for top human-level play.</p> |
| | </HTML> |
| | |
| | |
| | ==== Times for AI to cross human-relative ranges ==== |
| | |
| | |
| | <HTML> |
| | <p>Given the above dates, we have:</p> |
| | </HTML> |
| | |
| | |
| | <HTML> |
| | <figure class="wp-block-table"> |
| | <table> |
| | <tbody> |
| | <tr> |
| | <td>Range</td> |
| | <td>Start</td> |
| | <td>End</td> |
| | <td>Duration (years)</td> |
| | </tr> |
| | <tr> |
| | <td>First attempt to beginner level</td> |
| | <td>1968</td> |
| | <td><1987</td> |
| | <td><19</td> |
| | </tr> |
| | <tr> |
| | <td>Beginner to superhuman</td> |
| | <td><1987</td> |
| | <td>2017</td> |
| | <td>>30</td> |
| | </tr> |
| | <tr> |
| | <td>Above superhuman</td> |
| | <td>2017</td> |
| | <td>>2020</td> |
| | <td>>3</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </figure> |
| | </HTML> |
| | |
| | |
| | <HTML> |
| | <p><em><em>Primary author: Katja Grace</em></em></p> |
| | </HTML> |
| | |
| | |
| | ===== Notes ===== |
| | |
| | |
| | <HTML> |
| | <ol class="easy-footnotes-wrapper"> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-1-2680"></span>“Go Ranks and Ratings.” In <em>Wikipedia</em>, June 20, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&oldid=963489455">https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&oldid=963489455</a>.<a class="easy-footnote-to-top" href="#easy-footnote-1-2680"></a> |
| | </div></li> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-2-2680"></span>See table:<br/> |
| | <br/> |
| | “Go Ranks and Ratings.” In <em>Wikipedia</em>, June 20, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&oldid=963489455">https://en.wikipedia.org/w/index.php?title=Go_ranks_and_ratings&oldid=963489455</a>.<a class="easy-footnote-to-top" href="#easy-footnote-2-2680"></a> |
| | </div></li> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-3-2680"></span>“The first Go program was written by <a href="https://en.wikipedia.org/wiki/Albert_Lindsey_Zobrist">Albert Lindsey Zobrist</a> in 1968 as part of his thesis on <a href="https://en.wikipedia.org/wiki/Pattern_recognition">pattern recognition</a>.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-11">[11]</a></sup> It introduced an <a href="https://en.wikipedia.org/wiki/Influence_function_(statistics)">influence function</a> to estimate territory and <a href="https://en.wikipedia.org/wiki/Zobrist_hashing">Zobrist hashing</a> to detect <a href="https://en.wikipedia.org/wiki/Ko_rule">ko</a>.”<a class="easy-footnote-to-top" href="#easy-footnote-3-2680"></a> |
| | </div></li> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-4-2680"></span>“Figure 25 shows estimates from two<br/> |
| | sources: David Fotland—author of The Many Faces of Go, an Olympiad-winning Go program—and Sensei’s Library, a collaborative Go wiki. David Fotland warns that the data from before bots played on KGS is poor, as programs tended not to play in human tournaments and so failed to get ratings.” |
| | <p>Grace, Katja. “Algorithmic Progress in Six Domains.” Berkeley, CA: Machine Intelligence Research Institute, 2013.<a class="easy-footnote-to-top" href="#easy-footnote-4-2680"></a></p> |
| | </div></li> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-5-2680"></span>“In May 2017, AlphaGo beat <a href="https://en.wikipedia.org/wiki/Ke_Jie">Ke Jie</a>, who at the time was ranked top in the world,<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-27">[27]</a><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-28">[28]</a></sup> in a <a href="https://en.wikipedia.org/wiki/AlphaGo_versus_Ke_Jie">three-game match</a> during the <a href="https://en.wikipedia.org/wiki/Future_of_Go_Summit">Future of Go Summit</a>.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-wuzhensecond-29">[29]</a></sup>“<br/> |
| | <br/> |
| | “Computer Go.” In <em>Wikipedia</em>, July 27, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Computer_Go&oldid=969736537">https://en.wikipedia.org/w/index.php?title=Computer_Go&oldid=969736537</a><a class="easy-footnote-to-top" href="#easy-footnote-5-2680"></a> |
| | </div></li> |
| | <li><div class="li"> |
| | <span class="easy-footnote-margin-adjust" id="easy-footnote-bottom-6-2680"></span>“In October 2017, DeepMind revealed a new version of AlphaGo, trained only through self play, that had surpassed all previous versions, beating the Ke Jie version in 89 out of 100 games.<sup><a href="https://en.wikipedia.org/wiki/Computer_Go#cite_note-alphagozero-30">[30]</a></sup>“ |
| | <p>“Computer Go.” In <em>Wikipedia</em>, July 27, 2020. <a href="https://en.wikipedia.org/w/index.php?title=Computer_Go&oldid=969736537">https://en.wikipedia.org/w/index.php?title=Computer_Go&oldid=969736537</a><a class="easy-footnote-to-top" href="#easy-footnote-6-2680"></a></p> |
| | </div></li> |
| | </ol> |
| | </HTML> |
| | |
| | |
| |
