=Paper=
{{Paper
|id=Vol-2448/SSS19_Paper_Upload_219
|storemode=property
|title=A Study of Basis on AI-based Information Systems: The Case of Shogi AI System "Ponanza"
|pdfUrl=https://ceur-ws.org/Vol-2448/SSS19_Paper_Upload_219.pdf
|volume=Vol-2448
|authors=Yuichi Yoda,Kosuke Mizokushi,Seiichiro Honjo
|dblpUrl=https://dblp.org/rec/conf/aaaiss/YodaMH19
}}
==A Study of Basis on AI-based Information Systems: The Case of Shogi AI System "Ponanza" ==
https://ceur-ws.org/Vol-2448/SSS19_Paper_Upload_219.pdf
A Study of Basis on AI-based Information Systems:
The Case of Shogi AI System “Ponanza”
Yuichi YODA
US-Asia Technology Management Center, Stanford University
521 Memorial Way, Stanford, California, U.S.A
yyoda@stanford.edu
Kosuke MIZUKOSHI
Faculty of Economics and Business Administration, Tokyo Metropolitan University
1-1 Minamiosawa, Hachioji city, Tokyo, Japan
mizukoshi-kosuke@tmu.ac.jp
Seiichiro HONJO
Department of Management of Business Development, Shizuoka University
3-5-1 Jouhoku, Naka-ku Hamamatsu city, Shizuoka, Japan
seihonjo@shizuoka.ac.jp
Abstract consumption behavior. In contrast, Amazon’s recommen-
The objective of this study is to deepen our understanding dation system and Google’s search engine, which use ma-
about the exploration of Artificial Intelligence (AI) in cor- chine learning, respond to user needs based on data accu-
porate marketing and interpret how people and society re- mulated from the customers’ buying patterns. These sys-
spond in their attempt to comprehend the development and
tems do not rely on causal relationships and work as long
actions of AI.
In this study, we discuss the case of Ponanza, an AI as there is a correlation between data (Mayer-Schönberger
based system for Japanese Chess “Shogi”. We conclude that and Cukier, 2013). In marketing practice, there are three
Ponanza became a mystery even for its developers in their merits of focusing on the “result” of selections made by
process of building this system into one capable of defeating users in the real environment and responding to user needs
professional Shogi players and is now open to interpretation
through trial and error – (1) it conforms with the way of
for its developers and professionals.
In particular, when we treat AI as an extension of humans, thinking of companies that focus on results, (2) it leverages
it will be important to consider how AI and humans create the low cost of needs exploration in Internet business, and
knowledge and how humans can learn from AI. In the future, (3) it enables the company to incorporate complicated en-
interaction with AI can be expected to improve human’s vironmental factors which move dynamically when a pro-
ability to investigate “causes” and develop “reasons”.
posal is made to the user (Yoda, Mizukoshi and Honjo,
2016). This method is expected to grow in future.
Introduction On the other hand, such result-focused practice is disso-
ciated with existing human activity from the viewpoint of
With the spread of AI in recent years, there has been a systematic understanding because it does not specify cause
change in the way people understand its occupational prac- and effect. From the perspective of research, this leads to
tice. Many of the occupations currently performed by hu- difficulties in constructing theories on user needs or human
mans are expected to be replaced by AI in the near future behavior in general. From the practical perspective, this
(Frey and Osborne, 2017). Not only that, but the way oc- raises difficulties because the “results” cannot be repro-
cupational practices are undertaken is also undergoing duced as they are limited to certain conditions and horizon-
change. Until now, marketing research placed importance tal expansion of business is not easy. This study focuses on
on building hypotheses and providing reasons about users’
AI that surpasses human achievements to analyze the large number of situations on the board, calculates all pos-
workings of human understanding for a phenomenon and sible moves from those situations and predicts how the
the relationship between AI and humans. game is likely to unfold. In the case of humans, this is
called “reading” which means exploring (Yamamoto, 2017,
Ch. 1, Sec. 3, Para.8). However, because it is difficult to
Research Method completely explore all of the large number of situations
We study the case of a Shogi program that uses AI (Shogi due to resource constraints, computers determine the next
AI). Shogi is the best subject for studying the relationship move while marking out some highlights. This process of
between AI and humans because of the following three marking highlights is referred to as evaluation. In other
reasons - (1) it is a game with fixed rules, played in a static words, the area of exploration is gradually reduced as
environment and can therefore be studied as a case separat- needed to effectively use the limited resources (Ibid., Ch. 1,
ed from the complicated and dynamic environment of soci- Sec. 3, Para.10-12). Yamamoto said that humans program
ety, (2) professional Shogi players are considered one of the “exploration” part, which was a main function, and
the representative examples of the human intellect and (3) specified how the exploration was to be conducted, while
superiority dispute between AI and humans is already set- the computer learns to “evaluate” by itself through the in-
tled with Shogi AI far surpassing Shogi players. troduction of machine learning (Ibid., Ch. 1, Sec. 14, Pa-
In this study, the “case study” method as a qualitative re- ra.1).
search is adopted. Case study is an effective method for
exploratory research allowing us to ask “how” and ”why” Table 1 Major matches between Shogi AI and Shogi players
Year Details
of high-context phenomena beyond the control of the re-
Exhibition match between Bonanza and Akira
searcher (Yin, 1994)[4]. It is also suitable for exploratory Watanabe, Ryuou (Winner)
research of unique cases. The case study approach can be 2007
conducted adhering three principles of data collection to Bonanza made open source *partially used as
handle qualitative data as a scientific research approach reference for Ponanza too
proposed, vis-à-vis data correctness such as (1) use of mul-
2012 Bonkras (Winner) vs. Kunio Yonenaga, Eisei Kisei
tiple sources, (2) use of face-to-face interview, (3) mainte-
nance of a chain of evidence by Yin (1994). Ponanza (Winner) vs. Shinichi Satoh, 4-dan
This case study is based on an interview with Issei 2013 *Shogi AI’s first victory over an active profes-
Yamamoto1, the developer of major Shogi AI Ponanza, his sional Shogi player
public lectures, books and other related disclosed materials. 2014 Ponanza (Winner) vs. Nobuyuki Yashiki, 9-dan
Additionally, we requested Seiya Tomita (3-dan player in
the Encouragement Meeting) of the Japan Shogi Associa- Ponanza (Winner) vs. Yasuaki Murayama, 9-dan
tion to accompany us during Issei Yamamoto’s interview 2015 Winner of the 25th World Computer Shogi Cham-
and lectures. Before and after the events, we benefitted pionship
from his expert knowledge on the thinking process in-
volved in Shogi. We also used interview videos and books 2016 Ponanza (Winner) vs. Takayuki Yamasaki, 8-dan
by Shogi players Yoshiharu Habu and Amahiko Satoh as
Ponanza (Winner) vs. Amahiko Satoh, Meijin
reference materials for the analysis. 2017
*Shogi AI’s first victory over an active Meijin
Case Study of Shogi AI Ponanza2
(1) Overview of Ponanza It can be said that the difficulty in evaluating Shogi lies
Ponanza is a Shogi program that Issei Yamamoto started in the fact that no optimum method of calculation has been
developing while he was studying in the University of To- found for computers yet because of the game’s complexity
kyo. As Shogi AI, it defeated a professional Shogi player and depth4. Ponanza needed a function to express the ad-
for the first time on March 30, 2013. On May 20, 2017, it justments between the more than one hundred million pa-
became the first Shogi AI to beat an active Shogi “Meijin” rameters as “evaluation parameters” in order to represent
which is the most prestigious title of Shogi in Japan. the complexity of Shogi, based on three-piece relationships,
Yamamoto explained that as in the case of human including the king5, and the turns6. Yamamoto said that the
intellectual activities, Shogi AI too requires two functions initial version of Ponanza improved Shogi AI to a level
– exploration and evaluation. Exploration, here, refers to where it could play moves similar to about 45% of the pro-
the ability to predict and correctly emulate (make a guess fessional Shogi players 7 . This was done by enabling the
without adding one’s subjective views or judgement) the computer to adjust the values after using machine learning
future3. To anticipate the future, the computer explores a to acquire the game records of over 50,000 Shogi players
as training data. Machine learning based parameter adjust- ingly, even experts find it difficult to explain why random-
ments by computers are faster and more accurate than ly shared methods work well. They say that their best pos-
manual parameter adjustments by humans. Therefore, sible explanation is that “an experiment turned out well.”
Yamamoto decided to thoroughly train (adjust parameters) To sum up, Yamamoto says that he is unable to provide
the computer for the parameter function and devoted him- an accurate explanation of why Ponanza is strong and adds
self to describing through a program how the computer that he can only make it stronger through experimentation
should be trained to evaluate. and experience.
Then, on March 30, 2013, the computer defeated an pro-
fessional Shogi player for the first time. The match was (3) Shogi Players’ Perspective
played against Shinichi Satoh, 4-dan, in the 2nd Den-o Sen In the 2nd Den-o Sen, held in April and May 2017, Ponan-
(Electronic King Championship). The Ponanza at that time za became the first Shogi AI to defeat an active Meijin.
was able to explore 40 million situations in one second. Yamamoto mentioned in the 53rd turn of the first game
Furthermore, Yamamoto introduced reinforcement as an example of symbolic moves by Shogi AI. In this,
learning, which is unsupervised learning, in 2014, after Shogi AI made an exceptional move to build defense by
working on supervised learning where Ponanza learned giving up a piece to the Meijin who had no attacking piec-
from game records of Shogi players8. In this method, the es. This was against Shogi theory, in which players are
computer makes speculative searches even if the environ- expected to move across the board without giving up piec-
ment is unfamiliar and learns through feedbacks received es to the attacking player.
about the results. Repeated feedbacks strengthen the com- The opponent, Satoh Meijin, too felt surprised that such
puter’s evaluation function. To be specific, it makes six to a move was possible when he saw this happen. He said he
eight moves based on a probable situation, analyzes was unable to understand the meaning of the move because
whether they led to victories and finetunes evaluation pa- he held preconceived notions such as sacrificing a pawn
rameters. Yamamoto says that he accumulates about eight when the opponent has two pieces that are effective. He
billion such situations and has eventually analyzed nearly said that he was unable to anticipate the move. (Satoh Mei-
one trillion situations. This process results in determining jin, NHK, July 31, 2017). Moreover, Yoshiharu Habu ex-
new Ponanza-style tactics, which refers to sequences that plained “humans found it difficult to imagine a situation
do not exist in games played between humans. where a player would use a piece that is neither attacking
nor defending, and even give up a pawn to the opponent
(2) Developer’s Perspective who does not have one”. (Habu, NHK 2017).
As Ponanza’s performance improves, it is becoming more Satoh Meijin says that this showed that there could be
and more difficult to be explained. Yamamoto compared best moves in Shogi that humans do not see any reason for
its mystery to “Black magic9.” This term is accepted as a (that humans find difficult to understand). (Satoh Meijin,
slang in the machine learning world too and refers to an NHK 2017).
umbrella term for any technique whose origin and reason Shogi AI has already surpassed Shogi players. The dis-
for effectiveness is unknown. pute of superiority between humans and AI has been set-
When making improvements in Ponanza, every time tled as far as Shogi is concerned. However, humans, in-
Yamamoto thought of a new improvement, he would initi- cluding Shogi players, have not abandoned Shogi. Shogi
ate about 3,000 automatic matches between the Ponanza on players are beginning to find ways to learn from Shogi AI
which the improvement was applied and an older version. as part of their research on the game. Tomita, who fre-
The improvement would be implemented if the new quently participates in study groups with Shogi players,
Ponanza won 52% or more matches. However, Yamamoto says that the Shogi players he knows are placing im-
says that he had no clue about the workings of the im- portance on learning positioning judgement from Shogi AI.
provements that proved effective. In concrete terms, he In concrete terms, this means that Shogi players can im-
says that he does not understand the real reason why the prove their game by comparing their evaluation results
values fed in the program work or why a certain combina- with those of Shogi AI and refining their positioning
tion of values is effective. Yamamoto adds that he cannot judgement for each situation on the board.
analyze Ponanza’s effectiveness because he does not know
why it wins or loses a match, as the program’s Shogi
strength surpasses that of its developer Yamamoto. Discussion
As a concrete example of the black magic, Yamamoto It can be said that human understanding for the logic be-
gives the example of idle parallelization. In this method, hind Shogi AI’s strength has transformed through the fol-
multiple cores of the CPU separately carry out the same lowing three stages. In the first stage, players thought that
processing and the effective methods that each core acci- the logic behind Shogi’s strength lies within professional
dently discovers are shared with the entire system. Interest-
Shogi players and they sought to know how Shogi AI rep- can intervene in its contents. However, the difficulty of
licates the moves made by professional Shogi players. The determining the cause has become extremely increased and
focal point was whether the sources of strength that they it is a big challenge for humans.
expressed in their own language could be translated into On the other hand, in the approach where AI is consid-
machine language. In the second stage, the logic of ered an extension of humans, humans advance their under-
strength moved from professional Shogi players to Shogi standing of the results presented by AI while building a
AI. The developer of Bonanza discovered a method of con- model and seek “reasons” as a foundation for understand-
sidering the positioning of three pieces, rather than two, to ing. This corresponds with acknowledging as given the
determine the best move. At that time, it could be said that third stage, in which the logic of strength is beyond human
the translation was realized. The developer worked actively comprehension. As is evident from the comments of Shogi
on the logic of strength which had now moved to Shogi AI players, such as humans find it difficult to imagine a situa-
and attempted to adjust parameters. In the third stage, the tion where a player would use a piece that is neither attack-
logic of strength was internalized in Shogi AI and came to ing, nor defending, and even give up a pawn to the oppo-
be considered as something humans are unable to see. nent who does not have one (Habu and NHK, 2017), they
Shogi AI became capable of adjusting the logic of strength interpret AI as a person and try to learn from it. In such a
by itself through reinforcement learning. It was no longer case, humans and society, not AI, ask the foundation as to
clear why Shogi AI performed the way it did, and the de- “why” AI is able to produce certain results. That is why,
veloper’s adjustments came to be considered as black mag- contextual and language-based “reasons” should be ex-
ic. pected by humans and society. The challenge here is to
The internalization of the logic of strength in Shogi AI produce a logic to justify the foundation just like logical
meant that Shogi AI is superior to human Shogi players. At reasoning rather than discovering a working principle or
the same time, this also increased initiatives among profes- “cause.”
sional Shogi players to learn from Shogi AI. To learn from Various experiments in social psychology have shown
Shogi AI and improve their own game, professional Shogi that humans rationalize their actions using “reason” that
players must translate the logic of Shogi AI’s strength back are different from the “cause.” For example, consumers
into their own language. It will be important once again for were asked to explain the “reason” for selecting the most
professional Shogi players to be able to understand and high-quality nylon stockings from among identical prod-
explain why their Shogi AI opponent makes certain moves. ucts. Although a larger number of the consumers selected
Shogi players are tasked with learning and explaining stockings kept on the right side, they mentioned the differ-
why something worked, although they do not understand it. ence in the quality of the stockings while explaining their
This gives rise to the problem of how humans deal with a reason and not the positioning (Nisbett and Wilson, 1977).
phenomenon and gain an understanding of it as far as the Although AI does not automatically justify its actions by
relationship between humans and AI is concerned. By gen- itself, its actions can be seen as an extension of human be-
eralizing the example of Shogi AI, we find that humans are havior if we consider it as a subject for seeking “reason”
capable of taking two kinds of approaches. One is to treat after the action is performed.
AI as a physical phenomenon and the other is to deal with Moreover, when we consider AI as an extension of hu-
it as an extension to humans. mans, we can also anticipate the approach the other side
The first approach of treating AI as a physical phenome- takes when thinking. The Organizational Knowledge Crea-
non would mean that we just need to confirm that a certain tion Theory holds that knowledge is created, shared in the
result is produced under certain conditions, even though organization and accumulated through the repeated process
we do not understand the logic behind it. At this stage, it of four phases of the SECI model: Socialization, where
will be possible to apply AI using the phenomenon. More- people create or integrate tacit knowledge by sharing expe-
over, the current relationship between AI and humans can rience; Externalization, where people express tacit
be interpreted as similar to the relationship between the knowledge in clear concepts to convert it into explicit
steam engine and humans, before thermodynamics was knowledge; Combination, where people combine concepts
discovered. Humans learned from the steam engines, to build a knowledge structure; and Internalization, where
which was produced from experience; investigated the people embody explicit knowledge into tacit knowledge
causal relationship in its working principle and built the (Nonaka and Takeuchi, 1995). During externalization in
thermodynamics theory, thereby gaining a systematic un- the SECI model, tacit knowledge is converted into explicit
derstanding. This requires an investigation of the “cause” knowledge through dialogue between individuals. The
of a phenomena without depending on language. It can be concept of dialogue between two humans may be extended
said that from the third stage, where the logic of strength is to imagine an interaction between AI and humans where
internalized in Shogi AI, humans try to deepen their under- the latter learns from the former. As an extension to that, it
standing in order to return to the second stage where they might be effective to try to provide an explanation to this
highly persuasive phenomenon from the acts of personified
AI, using human abilities for intuition and logic. This professional “Meijin” for the first time.
2 This case study is based on Issei Yamamoto’s lecture (held at
could lead to new research questions, such as what the
meaning of a dialogue with AI is or whether there is a dif- Tokyo Metropolitan University Akihabara Campus) on September
ference between people who can learn from AI and those 22, 2017, a private interview with him on the same day and his
book.
who cannot. 3 Yamamoto, 2017, Ch. 1, Sec. 3, Para. 5
4 For instance, making computer-based calculations about the
quality of a situation is considered more difficult and tough to
Conclusion handle in Shogi than in Chess. In chess, the presence or absence
of pieces on the board is directly related to how good or bad the
In this study, we analyzed the case of Shogi AI to explore
phase is and can be represented more easily in the form of logic.
how humans comprehend the performance of AI that pro- However, in Shogi, the positioning of the pieces determines the
duces results that surpass humans. There are various im- quality of the situation and is therefore difficult to represent in the
portant social issues about AI that need to be handled. form of logic. (Yamamoto, 2017, Ch.1, Sec.10).
5 Yamamoto, 2017, Ch. 5, Sec. 1
These include who is responsible for the moral obligations
6 Retrieved February 20th,2018, “Ponanza Document (2010)
with respect to the results produced independently by AI
http://www.computer-
and how to handle social biases that are already a part of
shogi.org/wcsc20/appeal/Ponanza/Ponanza.pdf”
the training data. However, these issues are based on the 7 Yamamoto, 2017, Ch. 1, Sec. 12, Para. 7
premise of using AI as a tool governed by humans. This 8 Yamamoto, 2017, Ch. 1, Sec. 13
study sheds light on the approaches of treating AI as a 9 This refers to the magic used by witches to make mysterious
physical phenomenon and an extension of humans, and medicines in the world of fairy tales and fantasy. (Yamamoto,
shows that these approaches give rise to problems of a dif- 2017, Ch. 2, Sec. 1, Para. 2)
ferent kind. In particular, when we treat AI as an extension
of humans, it will be important to consider how AI and
humans create knowledge and how humans can learn from
References
AI. AI is producing better results than humans in various
fields. In the future, interaction with AI can be expected to Frey, C. B., and Osborne, M. A. 2017. The future of employment:
improve human’s ability to investigate “causes” and devel- how susceptible are jobs to computerisation?. Technological
Forecasting and Social Change, 114:254-280.
op “reasons.”
Habu Y. and NHK Special Coverage Groupe 2017. The core of
artificial intelligence:NHK Shuppan
Acknowledgements Ichinose M. 2006. Labyrinth of causes and reasons: Philosophy
of ‘Why’: Keisoshobou
In this study, we learned immensely from Issei Yamamoto, Mayer-Schönberger, V., and K. Cukier 2013. Big data: A revolu-
the developer of Shogi AI Ponanza, about his experiences tion that will transform how we live, work, and think:Houghton
and knowledge through the interview with him. We also Mifflin Harcourt.
received a wealth of information about the perspectives of Nisbett, R. E. and Timothy D. W. 1977. Telling more than
Shogi players from Seiya Tomita, a member of the Japan we can know: Verbal reports on mental processes, Psychological
Review 84(3):231-259.
Shogi Association. We are grateful to both for their invalu-
able support. Funding from the Telecommunication Ad- Nonaka, I. and H. Takeuchi 1995. The Knowledge-Creating
Company:New York: Oxford University Press.
vanced Foundation, Ritsumeikan University, Japan Mar-
Yamamoto I. 2017. How did AI exceed Shogi Meijin?: The es-
keting Academy, and JSPS KAKENHI Grant Number
sence of Machine learning, Deep leaning and Reinforcement
JP18K12878 is gratefully acknowledged. learning by the developer of AI Ponanza which is the strongest
Shogi AI:Diamond [Kindle version] Retrieved from Amazon.com
Yin, R. K. 1994. Case Study Research, Y Sage Publications, Inc.
Endnotes
Yoda Y., Mizukoshi K. and Honjo S. 2016. A Consideration of
the "Result" and "Reason" in the Exploration Process of the User
1 Issei Yamamoto, Developer of Shogi Program “Ponanza”, Spe- Needs utilizing AI : From Cases of Amazon. com and
Google :The Ritsumeikan business review, 55(3): 22-47.
cialty Appointed Associate Professor at Aichi Gakuin University,
Research Associate at Research Center for Advanced Science and
Technology of Tokyo University, Lead Engineer at HEROZ, He
started developing Shogi program Ponanza while he was a stu-
dent in the Faculty of Engineering, University of Tokyo. The
program became the first Shogi AI to defeat an active profession-
al Shogi player in a public match in the Shogi Den-o Sen event
held in 2013. In the 2017 Shogi Den-o Sen, it defeated an active