Online Match Prediction in Shogi Using Deep Convolutional Neural

Networks

Jim O’Connor and Melanie Fernández

Department of Computer Science, Connecticut College, New London, U.S.A.

Keywords:

Shogi, Deep Learning, Classiﬁcation, Evaluation Function, Dynamic Match Prediction.

Abstract:

This paper presents a novel approach to online evaluation of shogi games using Deep Convolutional Neural

Networks (DCNNs). Shogi, a complex deterministic abstract strategy game, poses unique challenges due to its

extensive game tree and the dynamic nature of piece movement, including the ability to play captured pieces.

Traditional methods of game evaluation for shogi rely on either expert knowledge and handcrafted heuristics,

or prohibitively high computational costs and limited scalability. Our method promotes a unique dataset

of shogi game records and SFEN (Forsyth Edward Notation) strings to convert board positions into binary

representations, which are then fed into a DCNN. The DCNN architecture, tailored for shogi board analysis,

consists of convolutional and fully connected layers culminating in a binary classiﬁcation output indicating a

winning or losing position. Training the DCNN on approximately one million board states resulted in an 82.7%

classiﬁcation accuracy on a validation set. Our approach allows for online single board evaluation, while

offering a computationally efﬁcient alternative to traditional methods, paving the way for the development

of additional shogi evaluation methods without the need for extensive expert knowledge or computational

resources.

1 INTRODUCTION

The development of intelligent agents to play de-

terministic abstract strategy games, such as shogi,

chess, and Go, has long been a cornerstone of ar-

tiﬁcial intelligence (AI) research. These games of-

fer an invaluable platform for probing the limits of

computational strategies and the development of ad-

vanced AI techniques. Among these games, shogi,

with its rich strategic depth and complex game me-

chanics, stands out as a particularly challenging do-

main. This complexity, coupled with the unique el-

ements of piece promotion and the ability to drop

captured pieces, underscores the need for innovative

approaches in AI game analysis and strategy devel-

opment. Historically, the quest to master strategic

games through computation has evolved dramatically,

from early heuristic-based engines to sophisticated

algorithms capable of achieving superhuman perfor-

mance. Beginning with pioneering work such as

Arthur Samuel’s ‘Samuel Checkers’ (Samuel, 1959)

and Tesauros ‘TD-Gammon’ (Tesauro, 1995), AI

agents were initially developed to lean heavily on

heuristics and simple learning techniques. These ini-

tial forays into the domain of game agents to con-

quer deterministic abstract strategy games were ex-

tremely successful, but relied in large part on the

simplicity of the particular game at hand. Conse-

quently, researchers were slow to expand this do-

main to more and more complex games. Some no-

table examples of this expansion were the conquering

of Western chess through the development of Deep-

Blue in 1997 (Campbell et al., 2002), and the more

recent success of AlphaGo in the game of Go (Sil-

ver et al., 2016). Although DeepBlue and other con-

temporary strategies relied heavily on hand-crafted

heuristics by expert players for evaluation functions,

the paradigm shift in game choice to the more com-

plex domains of Go and subsequently shogi have been

represented by a move into more complex computa-

tional approaches. This is clear in the current state of

the art programs like AlphaZero (Silver et al., 2018)

and MuZero (Schrittwieser et al., 2020), which em-

ploy deep reinforcement learning to master games of

chess, Go, and shogi. These systems demonstrate the

potential of AI to not only match but exceed the strate-

gic capabilities of the world’s best human players;

however, this efﬁcacy comes at a signiﬁcant cost of

extensive computational resources and training time.

600

O’Connor, J. and Fernández, M.

Online Match Prediction in Shogi Using Deep Convolutional Neural Networks.

DOI: 10.5220/0013018100003837

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Computational Intelligence (IJCCI 2024), pages 600-605

ISBN: 978-989-758-721-4; ISSN: 2184-3236

2 SHOGI

Shogi is a deterministic abstract strategy game, some-

times referred to as Japanese chess. In shogi, two

players compete on a 9x9 board (Figure 1) using

a number of pieces that can be easily compared to

chess and chess-variant counterparts; such as knights,

pawns, and kings. Some amount of strategic depth

is added through pieces such as the gold and sil-

ver generals, which can move in directions that are

unique from existing chess pieces. The most mean-

ingful rules difference in shogi for most contexts, is

the ’drop’ rule. When a piece is captured by an op-

posing player, that piece is then held ’in hand’ by the

player who captured it. On any player’s turn, they

may ’drop’ a piece that they hold in hand instead of

moving a piece. This recycling of the total pieces on

the board dramatically expands the tree of possible

game moves compared to other similar games, such

as chess. This signiﬁcantly larger search space is one

of the primary motivations for research using shogi as

a platform for AI agents.

Figure 1: A game of shogi in the starting position. Shogi

is played on a 9x9 board where each player starts with 20

pieces. Enemy pieces that are captured can be played again

by the capturing player on their turn.

The game of shogi itself has had a long and com-

plex history as a goal and domain for artiﬁcial in-

telligence research. Although the ﬁeld of computer

shogi has existed for decades, it wasn’t until land-

mark research and development of Bonanza in 2007

(Takizawa et al., 2015) that a program was able to

play shogi at a competitive professional level. Fur-

ther advancements in computer shogi continued with

the development of various additional heuristic-based

methods such as those developed by Wan (Wan and

Kaneko, 2018) and Grimbergen (Grimbergen, 1997),

which chipped away at the abilities of the top profes-

sional players throughout the next decade. Although

there was a signiﬁcant amount of success and positive

development during this time, these heuristic-based

methods required extreme amounts of hand-tuning by

domain area experts; only a handful of skilled devel-

opers in the world were able to contribute compet-

itive and consistent results in their engines. To the

contrary, later groundbreaking results in the develop-

ment of a computer shogi agent came through Ope-

nAI’s innovative work on AlphaZero and MuZero,

two agents that learn through self play and Deep Rein-

forcement Learning. These agents are able to achieve

absolutely superhuman levels of performance without

relying on handcrafted heuristics or expert level in-

volvement. However, the computational and temporal

costs associated with this learning technique are pro-

hibitive. Just as the expertise-based costs of the pre-

vious approaches left a competitive agent out of reach

for all but a few researchers and programmers, the

new costs in compute and running time in the state of

the art algorithms by OpenAI have once again left the

development of a competitive agent out of reach for

all but a select few. Our work focuses on a tractable

and necessary part of the game playing ecosystem via

single-board match prediction, and showcases a sim-

ple and effective approach with very low costs utiliz-

ing a novel dataset and the classifying power of Deep

Convolutional Neural Networks.

2.1 Match Prediction in Computer

Shogi

Within the ﬁeld of AI for computer game playing,

prediction of match outcomes is generally handled

in gestalt via an evaluation function. An evaluation

function is generally designed for each speciﬁc game

or task to act as a heuristic, offering an estimation of a

state’s ’quality’ or ’worth’ in a given problem domain.

In games such as shogi, AI agents begin by mapping

out a game tree with feasible moves through the use of

established algorithms like minimax or Monte Carlo

tree search. These algorithms then leverage the eval-

uation function to gauge the potential success of each

move within the context of the created game tree. The

effectiveness and precision of an evaluation function

are crucial for the performance of AI algorithms, es-

pecially in making decisions.

In the game of shogi, the role of the evaluation

function in AI algorithm development is critical due

to the game’s intricate strategies, the expansive 9x9

Online Match Prediction in Shogi Using Deep Convolutional Neural Networks

601

board, and the distinctive mechanism of reusing cap-

tured pieces. Analyzing a board state in shogi in-

volves considering various complex attributes like

material balance, the mobility of pieces, the safety of

the King, control over crucial squares, and the ability

to reintroduce captured pieces onto the board. These

elements together inform the score that the evaluation

function assigns to a board setup. Crafting an eval-

uation function demands thorough knowledge of the

speciﬁc problem area and often entails a trade-off be-

tween simplicity for computational speed and accu-

racy for optimal estimations. Although a traditional

evaluation function crafted in this way can be effec-

tive for game-playing agents, the signiﬁcant overhead

inherent to the evaluation being tightly coupled to the

agent as well as the expert knowledge necessary to

craft a competent evaluation function preclude their

efﬁcacy as a simple method of match prediction.

The latest advancements in shogi AI, exempliﬁed

by systems like AlphaZero and MuZero from Deep-

Mind, mark a departure from traditional, manually

crafted evaluation functions towards those generated

through self-play and reinforcement learning. These

systems learn exclusively from self-play, devoid of

human input, enabling the AI to discover a wide range

of strategies and tactics, some of which might be

unconventional yet highly effective. The evaluation

functions in these models are the result of deep neu-

ral networks trained across countless games of self-

play. This strategy has been extraordinarily success-

ful, propelling these AI to surpass human capabilities

not just in shogi, but also in other complex games like

chess and Go. However, the computational demands

of these innovative approaches have not only signif-

icantly raised the barrier to entry in the forefront of

computer shogi, but also present a meaningful and un-

necessary overhead to online match prediction.

3 METHODOLOGY

In the past few years, DCNN’s (Deep Convolu-

tional Neural Networks) have been commonly uti-

lized for tasks such as image classiﬁcation, object

detection, and semantic segmentation in matrices.

DCNN’s leverage multiple layers of non-linear pro-

cessing units for both low-level and high-level infor-

mation processing. This Deep Learning model is a

feedforward network that can be broken down into

two stages: feature learning and classiﬁcation. The

feature learning stage of the network consists of con-

volutional and pooling layers that are grouped into

modules and repeated according to the chosen ar-

chitecture. The convolutional layers extract features

from the input layer with weighted kernels and non-

linear activation functions that send the outputs to the

next layer. The objective of the pooling layers is to re-

duce spatial resolution and therefore to achieve spatial

invariance. The classiﬁcation stage of the network is

made up of a number of fully connected layers, end-

ing in a ﬁnal softmax or equivalent function providing

a classiﬁcation value.

Figure 2: An example of how a shogi board state is con-

verted to our binary matrix representation. A board state

from a professional game is pulled randomly from our novel

dataset. This board is then converted to SFEN, which is a

standard shogi board notation. This SFEN string is then

converted to a binary representation by assigning each let-

ter to a 5 bit binary string. The binary string is then shaped

into a 45x12 binary matrix, which is shown in this ﬁgure

using a black and white grid.

A large number of signiﬁcant advancements have

been made in the ﬁeld of AI through variations of

this architecture. Some notable examples, such as

ResNet (Targ et al., 2016) and Omnivec (Srivastava

and Sharma, 2023), can rival or even surpass human

level performance in image recognition tasks. Al-

though the game of shogi does not directly present

itself as an image classiﬁcation problem, we present

NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications

602

Figure 3: Our Deep Convolutional Neural Network architecture, which consists of a 45x12 binary input layer, a convolutional

layer with a 3x3 kernel and 128 output channels, and 4 fully connected layers, the last of which is a single output neuron. This

network classiﬁes our binary board state representation into a single ’black will win’ or ’white will win’ binary classiﬁcation.

a novel approach to leveraging the power of DCNNs

to classify image as a method of evaluating a win-

ning vs losing board position in a game of shogi. This

approach hinges on both our novel dataset of shogi

games and our shogi board representation that is us-

able by a DCNN.

The adaptation of a DCNN for shogi evalua-

tion entails the conversion of textual representations

of board positions, denoted in SFEN (Forsyth Ed-

ward Notation) strings, into binary bit strings. SFEN

strings serve as a textual representation of board po-

sitions, consisting of three ﬁelds separated by spaces.

The ﬁrst ﬁeld describes the pieces on the board; each

piece is represented by a letter, and each stretch of

spaces laterally is represented by an integer. So the

subsection "ln1g5" would represent a row of the board

ﬁlled by a Lance, a Knight, one empty space, a Gold

General, and ﬁve more empty spaces. A forward slash

then indicates the beginning of the next row, and the

process repeats. The second ﬁeld is either "w" or

"b" indicating the player who is next to move. The

third and ﬁnal ﬁeld represents the pieces in hand, with

black’s pieces being represented by capital letters and

white’s by lowercase. Our approach begins with a

novel dataset composed of approximately one mil-

lion SFEN strings gathered from records of profes-

sional shogi games. We convert each SFEN string to

a binary representation, with each square of the board

represented by ﬁve bits, and each additional metadata

character of the SFEN string being represented by a

further binary number (Figure 2). In total all of the

information included in the SFEN string is converted

into a 540 digit binary string. To convert the SFEN

to binary, each piece character is assigned a corre-

sponding 5 bit value, from 00001 to 11110. Empty

spaces are represented by a zeroed ﬁve bit string, and

the black and white players are represented by either

11111 or 00000, respectively. This binary string, as a

tensor, is utilized as one input to the DCNN.

3.1 Deep Convolutional Neural

Network Architecture

The Deep Convolutional Neural Network (DCNN) ar-

chitecture employed in this research (Figure 3) is tai-

lored for the analysis of shogi board states, treating

them as 45x12 binary matrices akin to black and white

images. As shown in Figure 3, the network begins

with an input layer, where the existing 540 bit board

state representation is converted into a tensor and fed

to the convolutional layer. The convolutional layer

has a kernel size of 3x3 and 1 pixel of padding to en-

sure the output feature map has the same spatial di-

mensions as the input. 128 distinct ﬁlters are created

by the convolutional layer, which are then ﬂattened

into another tensor and fed to the four continuous

fully connected layers. These layers progressively re-

duce the dimensionality of the feature space, eventu-

ally ending in a layer made up of a single neuron with

64 inputs, which uses a sigmoid squashing function to

output our single binary classiﬁcation. Rectiﬁed Lin-

ear Unit (ReLU) activation functions follow all but the

ﬁnal layer, introducing necessary non-linearity to the

network.

The loss function we leverage is the standard Bi-

nary Cross-Entropy Loss for one output, given as:

l = −(y log(p) + (1 − y)log(1 − p)) (1)

Where y is our labeled winner or loser, and p is

the predicted probability of the label being 1. For our

optimizer we leverage Adaptive Moment Estimation,

or Adam, with a learning rate of 0.001. Adam, de-

veloped by Kingma and Ba (Kingma and Ba, 2014),

is a ﬁrst-order gradient-based method for efﬁcient

Online Match Prediction in Shogi Using Deep Convolutional Neural Networks

603

stochastic optimization. We’ve found that this opti-

mizer is particularly applicable to our problem due

to its effective handling of sparse data representations

like ours.

4 RESULTS

Our Deep Convolutional Neural Network was trained

for 20 epochs, a greater than 80% successful classi-

ﬁcation rate in all runs of the network, and a 82.7%

success rate in the best case. The training data was a

random 80% of 200,000 board states pulled randomly

from our novel data set, and the validation data was

made up of the remaining 20% of that 200,000. Our

greater than 80% success rate is notable in a number

of ways. Notably, this level of accuracy allows for far

more successful dynamic match prediction based on

a single board than in previous works. In the more

tractable domain of chess dynamic match prediction,

Masud et al. (Masud et al., 2015) achieved a success

rate of nearly 66% under similar conditions. Related

attempts at binary classiﬁcation have confronted sim-

pler problems, particularly when attempting to tackle

the difﬁcult domain of shogi. Grimbergen was able to

achieve a success rate greater than 80% but on the far

more manageable problem of whether or not the king

was in danger, rather than determining the predicted

winner of the entire game from a single board. Our

promising results indicate a signiﬁcant step forward

in deterministic board game classiﬁcation and repre-

sent a number of new opportunities for game playing

agents that can be created without the prohibitive cost

of standard evaluation functions seen in other state of

the art programs.

5 CONCLUSIONS

Our results display several meaningful steps forward

in the domain of classifying shogi board states and

evaluating the position of a player in more efﬁcient

ways than has been shown previously. Our strategy

of using DCNN-based classiﬁcation allows us to give

an accurate estimate towards the winner of a game

of shogi without any input from a subject matter ex-

pert, instead using an online match predictor. This

classiﬁcation method can also be implemented and

used by developers with little to no experience in the

domain, due to the algorithm being agnostic of any

game rules or heuristics. Additionally, we are able

to make these predictions with a small fraction of

the computational resources and temporal resources

of other large state of the art algorithms. With our re-

Figure 4: The resultant classiﬁcation accuracy of ten runs of

our DCNN classiﬁer. These ten runs were executed sequen-

tially with consistent parameters and random subsections of

our one million board state dataset, as determined by our

80/20 training schedule.

sults achieving over 80% of accuracy in predicting on-

line match outcomes, this contribution presents itself

as a reasonable alternative to classiﬁcation of single

board states in comparison to the high computation

time that other shogi engines propose. This efﬁciency

allows us to train this classiﬁer in a matter of minutes

to hours on a standard desktop computer or laptop.

Consequently, this sophisticated shogi analysis can be

accessible to a broader audience that vary within skill

levels. Potential research plans will explore further

optimizations and applications as well as extending

these techniques to other complex strategy games and

enhancing their educational and competitive use.

REFERENCES

Campbell, M., Hoane, A., and Hsu, F.-h. (2002). Deep blue.

Artiﬁcial Intelligence, 134(1–2):57–83.

Grimbergen, R. (1997). Pattern recognition for candidate

generation in the game of shogi. CiteSeerX. Accessed:

2024-02-28.

Kingma, D. P. and Ba, J. (2014). Adam: A method for

stochastic optimization.

Masud, M. M., Al-Shehhi, A., Al-Shamsi, E., Al-Hassani,

S., Al-Hamoudi, A., and Khan, L. (2015). Online

Prediction of Chess Match Result, page 525–537.

Springer International Publishing.

Samuel, A. L. (1959). Some studies in machine learning

using the game of checkers. IBM Journal of Research

and Development, 3(3):210–229.

Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan,

K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E.,

Hassabis, D., Graepel, T., Lillicrap, T., and Sil-

ver, D. (2020). Mastering atari, go, chess and

NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications

604

shogi by planning with a learned model. Nature,

588(7839):604–609.

Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L.,

van den Driessche, G., Schrittwieser, J., Antonoglou,

I., Panneershelvam, V., Lanctot, M., Dieleman, S.,

Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I.,

Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel,

T., and Hassabis, D. (2016). Mastering the game of

go with deep neural networks and tree search. Nature,

529(7587):484–489.

Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai,

M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D.,

Graepel, T., Lillicrap, T., Simonyan, K., and Hass-

abis, D. (2018). A general reinforcement learning

algorithm that masters chess, shogi, and go through

self-play. Science, 362(6419):1140–1144.

Srivastava, S. and Sharma, G. (2023). Omnivec: Learning

robust representations with cross modal sharing.

Takizawa, T., Ito, T., Hiraoka, T., and Hoki, K. (2015). Con-

temporary Computer Shogi, page 1–10. Springer In-

ternational Publishing.

Targ, S., Almeida, D., and Lyman, K. (2016). Resnet in

resnet: Generalizing residual architectures.

Tesauro, G. (1995). Temporal difference learning and td-

gammon. Communications of the ACM, 38(3):58–68.

Wan, S. and Kaneko, T. (2018). Heterogeneous Multi-

task Learning of Evaluation Functions for Chess and

Shogi, page 347–358. Springer International Publish-

ing.

Online Match Prediction in Shogi Using Deep Convolutional Neural Networks

605