Overview

codecentric go challenge 2016

No Comments

This Sunday, August 28th, the third codecentric go challenge is going to start. The challenge – organized by Prof. Ingo Althöfer of the University of Jena, Germany, and sponsored by codecentric – is a best-of-five match of a strong computer go program (also called “bot”) against a leading amateur player. Franz-Josef Dickhut, amateur 6 dan and eleven-time German go champion, won the first two codecentric go challenges. In 2014 he defeated Crazy Stone, in 2015 Zen, both times with 3:1.

But since the codecentric go challenge 2015, the world of computer go has experienced a dramatic revolution. In March 2016, AlphaGo defeated Lee Sedol – undoubtedly one of the strongest go players ever – by 4:1. Motivated by this success, other go programs have improved dramatically as well.

The codecentric go challenge 2016 will pit Lukas Krämer, amateur 6 dan and reigning German Go champion, against the Japanese program Zen.

The games will be played online on KGS: https://www.gokgs.com. Lukas Krämer will play as “miao”, Zen as “Zen19X”. All games will start at 6pm CEST. The results will be documented on go.codecentric.de where you can also replay the games.

Like every year, we had a few questions for the competitors before the start of the match.

Interview with Lukas Krämer (LK), translated from German:

1. What are your experiences with Go bots?

LK: Lately I’ve often played lightning games against bots. But most of them are on 3 dan level on KGS. I have played a couple of games against 6 dan bots, too. Because of these games, I have the feeling that I can win against them with a high probability. But this depends on the bot as well as on my motivation to win. Most of the time I do something else simultaneously while playing against a bot. This would not feel right when playing against a person.
Also, these were lightning games. A longer thinking time changes a lot of things, as you can go through a more thorough consideration process.

2. If you followed the progress in computer go: What developments do you consider especially noteworthy?

LK: The question about the developments in computer go is obviously obligatory for a yearly event. But it seems a little bit superfluous with respect to the developments and the attention AlphaGo got last year. In short: AlphaGo is an unbelievable game changer. It shifts our understanding of go and effects a lot of positive movement in the go community – triggered not least by the new attention. It may also find soon applications in areas of our daily life. One good example was the enormous reduction of power costs of Google because of the application of the systems of AlphaGo. What’s also noteworthy is the rapid improvement of all other bots in a very short time and the fact that there seem to be many more bots playing on KGS now.

As I assume that every reader of this article has at least basic information about AlphaGo, I find the question about the impact of AlphaGo on us more important.

Some observers of the Lee Sedol – AlphaGo match remembered the changes computers had on chess – one of the most closely related games to us, maybe the closest one. Professional chess, especially at the top 10 level, is dominated completely by the computer analysis of certain openings and positions. Also e-cheating became a topic difficult to address. Bobby Fischer, who experienced this change and one of the greatest chess legends of history, emphasized this as the worst deficit in today’s chess.

I am relatively certain that go as a game is too complex. In my opinion the human limitations to memorize would make this impossible. I believe it unlikely that one can memorize computer responses for all candidate moves in all variations even when one looks only at one board position. If you try it anyway, it would possibly have the contrary effect. One would play worse as one leaves one’s own preparation more quickly. After that the question is who is the better player and has the better understanding of the game. Foreign ideas, whose deeper meaning you do not understand, can then be rather an impediment.

E-cheating is another issue. This is a question of the financial effort if one intends to prevent it as in the world of chess. But also here a modified version of AlphaGo could represent a major part of the solution.

In spite of these issues, which have a slightly negative ring to them, I see AlphaGo as a huge chance. But we have to make use of it.

I like to concur with a comment of Michael Redmond, 9 dan professional, about the consequences of AlphaGo on the game of humans against humans. He said he hopes that AlphaGo – if it re-learns go without injected games – inspires human players to break conventions and play freely just as Go Seigen did.

I have another hope: That the go community makes use of this unprecedented worlwide and media attention. I know that at the top level in Europe, the EGF, things are happening to position oneself in relation to the public and to work with this new chance. But it would be nice if more were be tried on country level, too. A broad base of players is the most important thing to spread go. This has been tried in the past but never worked really well. It is my dream that the DGoB uses money and effort to employ someone part-time. In my opinion, this person should then work professionally, effectively and trans-nationally as an organ of the DGoB to spread go, use marketing methods and search for sponsors. I am convinced this can be successful, as many good ideas and approaches failed or were not even tried, because the DGoB is a 100% volunteer organization. And these people – rightly – have to use their time for their jobs, too.

3. Did your view on go change since AlphaGo proved that machines can beat even the strongest humans in go?

LK: My view on go as a game did not change fundamentally. Before the match AlphaGo – Lee Sedol I had the impression that the bot is underestimated. But with this point of view I was relatively alone. Thus AlphaGo has surely opened the eyes a little bit for some people – including Lee Sedol. More amazing about the whole affair was the speed with which everything happened. If AlphaGo became stronger still, I would be happy to learn from the ideas of the bot and hopefully gain a better understanding of the game on a more abstract level.

4. How did you prepare for the match against Zen?

LK: To be honest, I do not know against which version I will play. We did not mention this while we were discussing other regulations. But I assume I will play against version Zen19b, which was already active on KGS, or a similar version. I looked through the games of Zen19b and I am convinced of its strength. This version has played convincingly and constantly powerful against strong players, among others American professionals and another 9 dan on KGS. Thus I believe that I would rate my chances rather negatively in a lightning match. But the difference between the mentioned games and the match against me will be the time settings. I hope that a longer thinking time will prove advantageous, as in my experience I play significantly better in tournaments with such settings. Unfortunately I do not know how Zen reacts to this change in time settings, but I am very excited to find out.

5. How do you see your chances?

LK: I am not sure of my chances. On the one hand, unfortunately I never could play against such a strong version of Zen before. On the other hand it must be pointed out that Zen has become much more powerful compared to one year ago. I am generally very convinced of my go. But it is always something different to play against a bot compared to playing against a human. They understand the game in another way than humans and this can be irritating. I intend to approach this match like all the others, too: With trust in my abilities and hopefully calm and relaxed, too.

Interview with Hideki Kato (HK), representing Zen for the codecentric go challenge. The answers are slightly edited:

1. Just after the codecentric go challenge 2015 AlphaGo entered the stage and revolutionized computer go. What did the advent of AlphaGo mean to you personally and to Team Zen?

HK: In December 2014, Aja et. al. published a paper of the application of DCNN to computer go and I had some personal communications with him. So, AlphaGo’s policy network was not surprising. What surprised me was its value network, which most developers thought almost impossible to build. Actually, the value network was not perfect (as shown in game 4 of the AlphaGo – Lee Sedol match), but most surprisingly even top professionals couldn’t attack its weak points (except in game 4).

This was observed, however, already with Crazy Stone. On KGS, Zen had a better rank than Crazy Stone (as well as direct matches), but the results against professionals were reversed.

I couldn’t expect the win of AlphaGo because the top professonals were weaker, especially in the early stages, than expected, and AlphaGo’s improvement in the half year before the match against Lee Sedol was bigger than expected.

That’s all for me. For team DeepZen (not Zen), Yamato recovered his motivation to improve Zen as AlphaGo showed the way.

2. Last year you mentioned three areas where go programs still had room for improvement: combining Life&Death solvers with the Monte Carlo framework, associative memories and learning methods. According to your knowledge, did AlphaGo use any of these approaches to achieve this amazing power?

HK: No. I’m afraid go will go the same way as chess and shogi, i.e. reach or even go beyond the level of human players without inventing human-like solutions. From the viewpoint of an AI researcher (me!), this is very disappointing because the game of go is just an application of AI research.

3. Was there anything in the concept of AlphaGo that surprised you in particular?

HK: Needless to say, the value network. Also, the belief that it can be built.

4. In what areas did Zen improve since last year? How much stronger do you think Zen has become?

HK: Mainly in the early stage of the game and end games (or static/silent positions) due to the use of DCNN to guide the search (policy network).

On the improvement since last year’s match: I guess two stones on average, but still several weak points caused by the Monte Carlo simulation remain and Lukas has chances to win. Evidence: the best handcap was four stones against top professionals in 2015 and is two stones these days. See Nick’s page about human-computer play.

5. What are your plans for Zen in the future?

HK: Now I will not be able to control Zen due to the DeepZenGo project supported by Dwango (at least) until next March. It’s very, very hard for me to have any idea on the future of Zen now. Please note: The project was started by Yamato’s hope, and I’m continuing to support him because this is the objective of the team and the DeepZen project.

I’ll study, anyway, how to learn not only one move but also a sequence of moves to overcome a weak point of current Monte Carlo simulations next year. The current candidate is DRNN. Also, I’d like to study how to introduce top-down control of the MCTS framework in the future. The current candidate is automatic capturing of models of outside (of an AI) by using ANN or augumented FSM.

6. Do you see any possibility that AlphaGo will be challenged as the top go program – be it by Zen or any other Go program?

HK: No doubt. The top runner is believed to be Zen, but I strongly believe there are some others almost being ready but not in the open, because the technology used for AlphaGo is not so difficult to implement if rich :).

7. What version of Zen will play in the codecentric go challenge 2016 and on what hardware?

HK: The version will be 12.0 but that may change. The hardware will be a dual Xeon server with a GPU (nVidia GTX 960) but that may change again, as this match takes at most five weekends and improved versions will be developed.

8. How do you view your chances in the match against Lukas?

HK: Most likely 3-1 for Zen. A 3-0 is also possible.

Glossary:

EGF: European Go Federation
DGoB: Deutscher Go Bund (German Go Federation)
DeepZen: The name of the team and the project, started in 2009: Yoji Ojima and Hideki Kato
Zen: The program itself
DeepZenGo: A cooperation project between DeepZen and Dwango, intended to beat AlphaGo
Yamato: Yoji Ojima, the chief programmer of Zen
MCTS: Monte Carlo tree search
DCNN: deep convoluted neural network
DRNN: deep recurrent neural network
ANN: artificial neural network
FSM: finite state machine

Post by Dr. Raymond Georg Snatzke

Game Programming

codecentric go challenge 2015

Uncategorized

codecentric go challenge 2015

More content about Game Programming

Comment

Your email address will not be published. Required fields are marked *