This article was originally published in Tech Spider – Know the Column And the same name public number (WeChat id: techspider)
An undisclosed mysterious Master suddenly appeared to sweep the human Go master and took 60 consecutive victories. An incomplete list of defeated players: Coulee, Ke Jie, Chen Yaozhen, Fan Tingyu, Chang Yi, Shi Yue, Yan Ting, Tang Weixing, Jiang Weijie, Yan Jiayu, Zhou Ruiyang, Pu Tingyu, Yuan Zhen, Jiang Dongrun, Jin Zhixi, Park Yong-Xun, Iyama Yuta. . . .
I guess the DeepMind team must have seen Chinese martial arts novels. A mysterious killer killed the rivers and lakes in the black winds of the moon. The martial arts were full of blood and dignity. The killer appeared when various rumors were on the rise. Wonderful marketing. A top-notch team that knows how to engage in gimmicks is like a pair of IQs with IQs around you, a gold-plated, long-selling colleague, you hate it!
In fact, in addition to AlphaGo, in recent years there are some very good artificial intelligence Go programs, they are often used with human chess players, each has a win, but the promotion is far less than Google. Before the advent of DeepMind’s AlphaGo, the masters of the artificial Go software were basically occupied by countries such as France and Japan. They have common basic ideas and unique highlights. Among them, Facebook’s darkforest provides detailed papers and code, and other computer Go programs are more mysterious.
Facebook’s Dark Forest (Darkforest)
The Dark Forest is an artificial intelligence Go program completed by two Chinese researchers at Facebook AI Research, named after Liu Cixin’s Trilogy II: Dark Forest. Dr. Tian Yuandong himself said that he was modest (original). If DeepMind decided to open it immediately after defeating Fan Wei in October 2015, or he would drag on himself for a while and decided not to vote for ICLR and wait until 2016 for ICML, then It was destroyed and there was no residue left. If the scientific researcher’s illusoryness is put aside, the rapid changes in artificial intelligence can be seen.
Like AlphaGo, the Dark Forest is also based on a deep neural network model (12 layers), so it has similar advantages and disadvantages as AlphaGo. The advantages are no longer described, and the defects mainly exist in the choice of local strategies. To compensate for this shortcoming, Dark Forest also chose Monte Carlo tree search as a complement to deep networks.
The Dark Forest has undergone three versions of evolution. The first generation of darkforest embodies the advantage of the traditional Monk Carlo search. The second generation of darkforest2 reached the level of stable KGS 3d. After the reference to the algorithm of alphaGo, the author added Monte in the darkforest2. Carlo algorithm, developed the third generation of darkfmcts3, the performance has been further improved, almost reached the level of KGS 5d. In fact, compared to the large number of teams in DeepMind, there are only two people in the development team of Darkforest. It is not easy to achieve such performance.
Compared to AlphaGo’s strategy network, which only predicts the next move, Darkforest can predict the predictive long-term moves, including its own and opponent’s chess, which is its unique and powerful place. Personally, when AlphaGo and Li Shishi played chess, they sometimes came out with some useless “computer hands”. On the one hand, DCNN’s inherent drawbacks (Darkforest’s paper also mentioned similar problems), on the other hand, it may be because Lack of long-term planning. Another unique feature is that the Darkforest architecture uses multiple softmax outputs to enhance supervision during training. To increase the speed of convergence, Darkforest also uses the latest deep residual network ResNet.
What Darkforest lacks than AlphaGo is the valuation network that predicts the overall situation. This part is AlphaGo’s originality. Personal guessing is also one of the reasons why it has a good overall situation.
Industry conscience, the training source code of the dark forest is open source, please click on the Github link if you are interested. facebookresearch/darkforestGo
Zen is one of the most famous Go programs currently available. The first version was released in 2009 and the latest version was released in June 2016. Unlike the dark forest and AlphaGo team development, Zen is a separate Go program developed by Japanese programmer Yoji Ojima, and the hardware part is implemented by Kato Hideki. Zen’s inferior hardware devices (single Mac Pro 8 cores, and of course the latest mini cluster or GPU version) have won championships in various Go artificial intelligence competitions for many years. The Zen 19K2 is also the first machine to reach 9D.
CrazyStone was developed by French computer scientist Rémi Coulom and used the Monte Carlo search method to run on the Grid’5000 large-scale computing platform. CrazyStone was defeated by Ishida Yoshio in 4 in 2013, and was defeated by Yoshiki in 4 in 2014. The picture shows Rémi Coulom and Yida Jiji are in the process of confrontation.
In May 2016, the latest version of CrazyStone added a deep learning module to maximize performance and achieve KGS 7d level, with a 90% win rate for CrazyStone 2013 with no deep learning. Unbalance CorporationAlso launched a commercial version, for $ 80 you can have a professional player level full-time sparring!
MoGo from France and Taiwan
MoGo is a Go software developed by INRIA (French National Institute of Information and Automation) and a team from Taiwan. In Taiwan in 2009, MoGo defeated Zhou Junxun’s nine-segment on a 19*19 full-size board under the condition of being 7 sons. The technologies used by MoGo are: the Monte Carlo method, the tree search algorithm based on Armed Bandits (which was revolutionary at the time), and the high-performance computing cluster (running on the Grid’5000 large-scale computing platform like Crazy Stone). Recently, MoGo seems to have not updated one step.
South Korea’s DolBaram
As a country with considerable accumulation in the field of Go and smart technology, NuriGrim, a small company in South Korea, decided to challenge Japan and France to develop DolBaram. In the 2016 UEC Cup, DolBaram scored second place. In the limited information that is circulated, you can know that DolBaram, like AlphaGo, also uses deep learning methods.
It can be seen that the impact of AlphaGo’s success on the industry of artificial Go software is obvious, and the combination of deep learning and Monte Carlo has become mainstream. After Go, similar methods are more exciting in other applications that require complex decision-making, such as medical care, government decision-making, smart cities, and so on. It is human wisdom that creates artificial intelligence, and your uncle is your uncle.
This article is an original article, if you need to reprint, please contact the technology spider WeChat public number background.
Some images come from the Internet. If there is any infringement, please contact us to delete.