10.30h Keynote Lecture: Gian-Carlo Pascutto (Leela Town, Belgium): Trends and Observations in Computer Go
11.15h Discussion 作者: lu01 时间: 2018-7-14 16:05
还有人机配对比赛
16.45h – 19.30h Human and Computer Pair-Go Demonstration Game: Kozo Hayashi - Leela Zero vs Manja Marz - Leela Zero 作者: hred9D 时间: 2018-7-15 10:04
比利时人的AI智能程序和足球都是非常强的感觉！作者: lu01 时间: 2018-7-15 10:18
对作者: book 时间: 2018-7-19 12:15
感谢GCP的贡献 作者: lu01 时间: 2018-7-28 08:52
大会开幕了作者: lu01 时间: 2018-7-31 13:31 https://tieba.baidu.com/p/5812363537作者: lu01 时间: 2018-8-9 08:46
等公布ppt作者: lu01 时间: 2018-8-10 05:30
PPT公布， https://github.com/gcp/leela-zero/issues/1706作者: lu01 时间: 2018-8-10 05:31
请能访问google的转发一份到这里作者: Joe 时间: 2018-8-10 08:54
求gcp照片作者: lu01 时间: 2018-8-10 17:39
Trends and Observations in Computer Go
Gian-Carlo Pascutto, gcp@sjeng.org
Up until 2006, computer Go had progressed at a snail's pace, with the best programs being as
strong as the average amateur. However, the discovery of a new Monte Carlo Tree Search
method set things in motion,
and they would not slow down.
Advances in computer vision, neural networks, graphics hardware, reinforcement and
supervised learning, pattern scoring and tree search have propelled computer go players
convincingly past humans in the 12 years that followed. The time has come where everyone
with a personal computer can potentially have a professional lever sparring opponent and
analysis partner. We will discuss what is still missing to make the picture complete.
The initial breakthrough was made by the introduction of an efficient Monte Carlo Tree
Search method applicable to Go. For the first time, the strength of a Go program scaled with
hardware. The ad hoc basis of the algorithm was quickly improved by linking the discovery
with existing research on ... playing slot machines. In some ways, humanity was doomed at
that point: further, inevitable increases in the speed of hardware would continuously improve
the strength of the machines. In 1989 Feng-hsiung Hsu drew a graph plotting the speed vs
strength of existing chess computers, and made a prediction about what would be needed to
beat the world champion. He delivered with Deep Blue 7 years later. In 2006 the distance to
human Go champions was big enough that no-one would draw such a plot, but the end result
has turned out to be no different.
A second breakthrough was made by the re-introduction of neural networks into Go. Once a
poster child of AI in the 1980's, but later discarded, advances such as convolutional networks,
solutions to vanishing gradient problem, residual stacks and optimizations such as Batch
Normalization made them powerful imagine recognition techniques, and advances in
graphics hardware quickly made very large networks practical. Multiple teams quickly found
out that their performance was just as good in Go, and a second jump in performance was
achieved, propelling us more quickly to the end goal.
The last jump came from improved reinforcement learning methods. With the programs
within striking distance, a last step needed to be made for the apprentice to surpass the
master. By finding improved methods to let the programs improve from their own mistakes,
computers finally surpassed the best humans.
Being at this point now, it is time to make the results usable to Go players. Leela Zero has
replicated the above research in such a way that the needed software and networks are
available to anyone. What do we need for good performance of a Go program, how do we
make use of it, what have we learned so far, and what parts are still missing or could be
improved? 作者: lu01 时间: 2018-8-10 17:41 https://www.uni-trier.de/index.php?id=65059作者: lu01 时间: 2018-8-11 20:17
[attach]148907[/attach]