3-х кратный Чемпион Европы Фань Хуэй проиграл AlphaGo 5:0!

[Dmitriy]

Як вам таке?:
http://www.usgo.org/news/2016/01/alphag ... i-advance/

[zivilizazia]

Як вам таке?:
http://www.usgo.org/news/2016/01/alphag ... i-advance/

не верю. похоже на фейк со стороны автора программы. пускай проиграют ещё другие, посмотрят другие авторы программ.. сколько уже рассказывали и обещали нейронщики, а воз и ныне там.

[Dmitriy]

PRESS RELEASE
MAJOR STEP FORWARD IN ARTIFICIAL INTELLIGENCE
A computer program developed by Google DeepMind (AlphaGo) to play the Oriental game of Go
has beaten the three-times European Go Champion and Chinese professional Fan Hui. This is the
first time that a Go-professional has lost such a match, and not only that, by a clean sweep in all 5
games. This signifies a major step forward in one of the great challenges in the development of
artificial intelligence - that of game-playing.
These findings were reported in a peer-reviewed study published in the scientific journal Nature:
Silver D. et al. Mastering the game of Go with deep neural networks and tree search. Volume 529,
issue 7587, pp 484-489: http://www.nature.com/nature/journal/v5 ... 16961.html.
The British Go Association recognises the achievement by Google DeepMind. We congratulate
them and look forward to the challenge it has issued to the top player in the world for the last 10
years, Lee Sedol from South Korea, to take place in Seoul in March.
Although we were not officially involved in the match between AlphaGo and Fan Hui, Toby
Manning, our Treasurer, was the Referee and observed the games closely. He says
"The games were played under full tournament conditions, and there was no disadvantage to
Fan Hui in playing a machine not a man. Google DeepMind are to be congratulated in
developing this impressive piece of software".
Our President, Jon Diamond, was one of the early researchers in Computer Go at London
University some 40 years ago (see http://britgo.org/computergo/history for some of the history) and
says:
"Following the Chess match between Gary Kasparov and IBM's Deep Blue in 1996 the goal
of some Artificial Intelligence researchers to beat the top human Go players was an
outstanding challenge, perhaps the most difficult one in the realm of games. It's always been
acknowledged that the higher branching factor in Go compared to Chess and the higher
number of moves in a game made programming Go an order of magnitude more difficult.
On reviewing the games against Fan Hui I was very impressed by AlphaGo's strength and
actually found it difficult to decide which side was the computer, when I had no prior
knowledge. Before this match the best computer programs were not as good as the top
amateur players and I was still expecting that it would be at least 5-10 years before a
program would be able to beat the top human players; now it looks like this may be
imminent. The proposed Challenge may well be that day.
One significant aspect of this match was that AlphaGo analysed orders of magnitude fewer
positions than Deep Blue did. DeepBlue also had a handcrafted evaluation function, which
AlphaGo does not. These indicate the general improvements in AI techniques that Google
DeepMind have achieved. This surely means that the technology behind it will be really
useful in other knowledge domains."
1
Hajin Lee, Secretary General of the International Go Federation and a Korean Go Professional,
said:
"AlphaGo's strength is truly impressive! I was surprised enough when I heard Fan Hui lost,
but it feels more real to see the game records.
My overall impression was that AlphaGo seemed stronger than Fan, but I couldn't tell by
how much. I still doubt that it's strong enough to play the world's top professionals, but
maybe it becomes stronger when it faces a stronger opponent."
An article by Toby Manning about the match, including the game records and comments, will
appear shortly in the next British Go Journal (the magazine of the British Go Association). Because
of the international significance of this match, we're publishing a publicly available version of this
article online on our website (http://www.britgo.org/deepmind2016).
Regarding the challenge match Lee Sedol said:
"This is the first time that a computer has challenged a top human pro in an even game, and I
am privileged to be the one to play it. Regardless of the result, it will be a meaningful event
in Baduk (Go) history. I have heard that Google DeepMind's AI is surprisingly strong and
getting stronger, but I am confident that I can win at least this time."
Jaeho Yang (Secretary General, Korean Baduk Association) said:
"We welcome the development of computer AI with Google DeepMind and its challenge to
Lee Sedol. It will be an exciting event for all Baduk players around the world. About the
result, I think the computer may win one or two games because Lee Sedol hasn't played any
strong computer AI before, but I doubt that the computer can win the overall match."
Just as the Kasparov/Deep Blue match did not signal the end of Chess between humans, so this
development of AlphaGo does not signal the end of playing Go between humans. Computers have
changed the way that players study and play Chess (see this Wired article from 2012:
http://www.wired.com/2012/09/usa-chess- ... -computer/), and we expect something similar
to occur in the field of Go, but not necessarily as assistance during play.
It has been recognised for a long time that achievements in game-playing have contributed to
developments in other areas, with the game of Go being the pinnacle of perfect knowledge games.
The previous major breakthrough in computer Go, the introduction of Monte-Carlo tree search, led
to corresponding advances in many other domains such as planning and constraint satisfaction.
Facial recognition seems just one of the obvious areas that the techniques developed by Google
DeepMind will be useful for.
------------------------------------------------------------------------------
INFORMATION FOR EDITORS
Go is one of the world's great strategy board games and originated over 3000 years ago in China. It
is part of popular culture today in Japan, China and Korea, and known as Igo, Weiqi or Baduk
respectively in these countries. It was brought to the UK nearly 100 years ago from Japan and
continues to expand in popularity.
There are about 1000 full-time professional players, almost all in Asia, with the top one probably
being Chinese. More than 10% of Koreans play the game and watch it on dedicated TV channels.
Go is a game of simple concepts - starting with an empty board, alternately placing pieces which
don't move (called stones) on it, trying to surround empty areas, whilst capturing your opponent's
stones by surrounding them and taking them off the board. However, as with a lot of simple
2
concepts, it can take a little while to grasp them and a long time to master them. Players are of all
ages, from 4 or even younger to more than 90!
One of its best features is the handicapping system. This allows players of very different strengths
to play a proper game giving each player a 50% chance of winning.
For more details about Go see http://www.britgo.org/whatisgo
------------------------------------------------------------------------------
The British Go Association is a voluntary organisation promoting Go in the UK, founded in the
mid-1950s.
We introduce Go to young people and others and assist with the development of the more than 60
clubs in the UK, with teaching events and seminars around the country.
We run an on-line team league and support the more than 30 tournaments a year all around the
country. We also select UK representatives for international events in Europe and Asia.
See http://www.britgo.org for full details.
CONTACT:
Jon Diamond (President)
Email: president@britgo.org
Phone: 01892 663837 (preferably) or 07885 372605
------------------------------------------------------------------------------
Google DeepMind is a British artificial intelligence company. Founded in 2011 as DeepMind
Technologies by Demis Hassabis, Shane Legg and Mustafa Suleyman, it was acquired by Google in
2014.
See http://deepmind.com
CONTACT
deepmind-press@google.com

[Di]

Партии AlphaGo (Google DeepMind) vs Fan Hui 2p результат 5-0

[zivilizazia]

PRESS RELEASE
MAJOR STEP FORWARD IN ARTIFICIAL INTELLIGENCE

Regarding the challenge match Lee Sedol said:
"This is the first time that a computer has challenged a top human pro in an even game, and I
am privileged to be the one to play it. Regardless of the result, it will be a meaningful event
in Baduk (Go) history. I have heard that Google DeepMind's AI is surprisingly strong and
getting stronger, but I am confident that I can win at least this time."

я понимаю, что эпатаж и гугль (via filiale DeepMind) всё держит в секрете.
но сплетничают, что:

1. процессор-монстр (1000 процессоров, база данных= попугай, который поёт. как человек, но это же не то и не оттуда. это, как крем тем, кто не знает и не думает, что он так же прост, как майонез, подают, как панацею вечной молодости из парижа за 800 евро от лучших гуманнейших специалистов, не тестирующих на животных).

2. какой-то так у них новый алгоритм революционный, даже не с бинарным кодом. теоретически возможно и такое. но это же не ум. это, как говорить, что инопланетяне выкрадывают людей.

предпочту с ахами подождать до марса и увидеть, что AlphaGo лучше Lee Sedol. для меня это будет действительно фактом, что AlphaGo не подсказывали.

спрашивается тогда, почему они взялись именно за го, как приоритет, а не пошли показывать нового Баха, Байрона, в химию, не использовали мощность на написание программы, которая бы всё сделала вместо программистов и тд

"We built a system, AlphaGo, that combines an advanced tree search with deep neural networks. These neural networks take a description of the Go board as an input and process it through 12 different network layers containing millions of neuron-like connections. One neural network, the “policy network,” selects the next move to play. The other neural network, the “value network,” predicts the winner of the game."

[Shedal]

Ещё одна статья про победу DeepMind AlphaGo над Fan Hui'ем:

http://www.nature.com/news/google-ai-al ... NjM2Nzg5S0

Там есть видео, где разработчики AlphaGo рассказывают о себе и о том, как работает их продукт.
Даже вкратце рассказывают про правила игры.

Научная статья, описывающая алгоритмы AlphaGo.

UPD: Та же статья в более читабельном виде.

Упрощённая сводка алгоритмов:

* Разработчики создали стратегию обучения, которую назвали SL (supervised learning — контроллируемое обучение). Согласно этой стратегии, программа учится на основании 30 миллионов настоящих человеческих ходов и пытается им подражать в своих играх. Стратегия реализована в виде свёрточной нейронной сети, которая на вход принимает позицию на доске, а на выходе отдаёт распределение вероятностей разрешённых ходов.
* Автоматизировали процесс играния этой нейронной сети с самой собой — миллионы раз подряд. Вернее, играет сеть со случайно выбранной прошлой версией самой себя — все прошлые состояния нейронной сети сохраняются именно для этой цели. Для выигранных игр вероятности правильных ходов в сети увеличиваются при помощи расчёта градиентов вероятностей на каждом ходу партии. Эта стратегия обучения называется RL (reinforced learning — усиленное обучение). Полученная при помощи RL нейронная сеть выигрывает у изначально обученной SL-сети в 80% игр.
* После получения двух обученных таким образом нейронных сетей, алгоритм играющей программы выбирает ход следующим образом: выполняется поиск Монте-Карло на основании SL-сети с отсечением позиций при помощи функции, основанной на RL-сети. Интересно, что SL-сеть лучше справляется с изучением графов возможных вариантов, а RL-сеть — с оценкой позиций.

Как можно видеть из этого описания, стратегии обучения и алогритм выбора хода даже не заточены под специфику го, а значит, могут использоваться и в других сферах. Например, разработчики планируют создать автоматический анализатор медицинских изображений (МРТ, рентген и т.п.), а также применить этот алгоритм к метеорологическим прогнозам.

[zivilizazia]

many challenges remain in DeepMind’s goal of developing a generalized AI system. In particular, its programs cannot yet usefully transfer their learning about one system — such as Go — to new tasks; a feat that humans perform seamlessly. “We’ve no idea how to do that. Not yet,” Hassabis says.

Algorithms alone cannot successfully play it on any advanced level. It requires a massive database of stored games for the computer to search. By itself, the strongest computer cannot successfully calculate from elements, rules and principles of opening, middle game and ending to accomplish master-level chess play. And, artificial intelligence will never successfully replicate the human mind. Kurt Godel demonstrated that in the 1930's with his famous incompleteness proofs. He even proved that a perfectly consistent mathematical system capable of generating elementary arithmetic from a set of axioms and rules of inference can never prove its own consistency within the system. The result is that any self-programming machine (an autonomous robot) will necessarily generate fatal inconsistencies and fail to "recognize" its own validities. It cannot have self-reference of "mind" (whatever that would mean in a robot.) That is, it could never be a successfully reflective entity such as the human mind. It's fascinating stuff. The title: "On Formally Undecidable Propositions of Principia Mathematica and Related Systems." It's a very difficult proof. Easier is the book "Godels Proof" by Nagel and Newman.


Архитектура свёрточной нейронной сети по большей части пригодна только для распознавания изображений. Слишком много варьируемых параметров сети, непонятно, для какой задачи и вычислительной мощности какие нужны настройки. Так, к варьируемым параметрам можно отнести: количество слоев, размерность ядра свёртки для каждого из слоёв, количество ядер для каждого из слоёв, шаг сдвига ядра при обработке слоя, необходимость слоев субдискретизации, степень уменьшения ими размерности, функция по уменьшению размерности (выбор максимума, среднего и т. п.), передаточная функция нейронов, наличие и параметры выходной полносвязной нейросети на выходе свёрточной. Все эти параметры существенно влияют на результат, но выбираются исследователями эмпирически. Существует несколько выверенных и прекрасно работающих конфигураций сетей, но нет правил, по которым нужно делать выбор для новой задачи.

к числу тех, кто считает, что интеллент IA превзойдёт интеллект человека ещё не скоро, принадлежат такие личности, как: Stephen Hawking (astrophysicien britannique), Bill Gates (fondateur Microsoft), Elon Musk (PDG de Tesla) и Stuart Russell (специалист в IA).

[Shedal]

Ещё одна интересная статья, опубликованная на официальном блоге Гугл.

Хмм, не думал, что я такое когда-либо скажу, но теперь я жду игр АльфаГо с Ли Седолем в марте даже больше, чем рематча Кличко против Фьюри в мае

[unnamed]

к числу тех, кто считает, что интеллент IA превзойдёт интеллект человека ещё не скоро, принадлежат такие личности, как: Stephen Hawking (astrophysicien britannique), Bill Gates (fondateur Microsoft), Elon Musk (PDG de Tesla) и Stuart Russell (специалист в IA).

Тут надо еще уточнять какого именно человека они имели в виду.
Что касается го то после того как надмозг гугла был выпущен на свободу вполне естественно было что произойдет значительный скачок.
Уже сейчас можно говорить что во вселенной осталось всего порядка тысячи людей, которых робот не мог бы обыграть в го.

[Di]

Комментарии топов о AlphaGo: