See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/272820284

Analyzing Positional Play in Chess using Machine Learning

Technical Report · December 2014

DOI: 10.13140/2.1.3886.9283

CITATION READS
1 562

3 authors, including:

Rohit Mundra
Stanford University
5 PUBLICATIONS   4 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Mid-latitude All-sky-imager Network for Geophysical Observation View project

Plankton Classification Using Hybrid Convolutional Network-Random Forests Architectures View project

All content following this page was uploaded by Rohit Mundra on 27 February 2015.

The user has requested enhancement of the downloaded file.

Analyzing Positional Play in Chess using Machine Learning
Sameep Bagadia Pranav Jindal Rohit Mundra
sameepb@stanford.edu pranavj@stanford.edu rohitm92@stanford.edu
December 13, 2014

Abstract 2 DATASET
With increases in computational power and better algo- In this section, we describe our dataset followed by the
rithms, chess engines are able to explore game trees of preprocessing steps that were required to structure the
greater depths and thus have excelled at calculative play. data.
However, since the growth of the move search space is
exponential, even with increase in computational power
chess engines perform only marginally better year-to- 2.1 Source and Structure of Data
year. In this paper, we explore a novel technique for
We used board positions from 6200 games from an engine
assessing board positions using machine learning tech-
vs. engine tournament in 2006 [1]. This provided us with
niques that can be used to supplement and improve cur-
a total of 1,075,137 board positions in Portable Game
rent chess engines. We model chess board positions as
Notation (PGN) where a board position is a snapshot of
networks of interacting pieces and use supervised ma-
the game at a point in time. We parsed the PGN data
chine learning techniques to analyze positions and pre-
to a more convenient Forsyth–Edwards Notation (FEN)
dict outcomes in chess games.
for further processing.

1 INTRODUCTION 2.2 Data Selection Criteria

Chess has two broad approaches to game-play, tacti- Each board position also had an associated score cal-
cal and positional. Tactical play is the approach of culated by the engine. A board score is a measure of
calculating maneuvers and employing tactics that take the relative strength of white compared to black; thus,
advantage of short-term opportunities, while positional a positive value indicates that the chess engine’s evalu-
play is dominated by long-term maneuvers for advantage ation believes that white is in a stronger position while
and requires judgement more than calculations. a negative value indicates the converse. If a board score
is near zero but the game then goes on to end decisively,
Current generation chess engines predominantly we hypothesize that aspects of positional play were un-
employ tactical play and thus outplay top human play- accounted for by the chess engine. Thus, we consider
ers given their much superior computational abilities. only those board positions where the board scores are
Engines do so by searching game trees of depths typi- nearly zero but the game ends decisively. It is worth
cally between 20 and 30 moves and calculating a large noting we restrict our focus to engine vs. engine games
number of variations. However, human play is often a where we know that the successive moves are going to
combination of both, tactical and positional approaches, be tactically optimal. Thus, the board positions we con-
since humans have some intuition about which board sidered are those whose outcome can be attributed to
positions are intrinsically better than others. positional aspects with high confidence. The hypothesis
can be seen in Table 1.
In our project, we use machine learning to identify
Table 1: Cause of Outcome
elements of positional play that can be incorporated in
chess engines. We model chess board positions as net- Engine Score
works of interacting pieces and predict game outcomes Outcome
≈0 0 0
when the engine evaluates both sides to be of compa-
White win Positional Tactical Improbable
rable strength. Our findings indicate that we can make
Black win Positional Improbable Tactical
such predictions with reasonable accuracy and thus, this
Draw Variable Improbable Improbable
technique can be augmented with current chess engines
to improve their performance.
3 NETWORKS & FEATURES
To capture interactions between various pieces and loca-
tions on the board, we modeled the board as networks
with board squares and pieces as nodes, and the interac-
tions as edges with attributes. This approach is similar
to that seen in [2]. For each board position, we create
the following two networks:
1. Support Network: This is a network where each
node represents a piece and each directed edge rep-
resents that the first piece can reach the second
piece. If the edge is between pieces of the same
color, then it is attributed to be a defense edge (since
the first piece can defend the second piece). Con-
versely, the edge between pieces of different color is
attributed to be an attack edge. Figure 1 demon-
strates attack and defense edges on a chess board.
The corresponding network for this board position (a) A Chess Board Support Network
can be seen in Figure 2a.
2. Mobility Network: This is a bipartite network
from chess pieces to board squares. An edge from
a piece to a location represents that the piece can
reach that location on its next move. Mobility edges
on a chess board can also be seen in Figure 1. The
(b) Mobility Network for White
corresponding network can be seen in Figure 2b.
Figure 2: Network Structures for Board Positions

• Mobility around the king: Captures the concept of

king-safety in chess, it is desired to have low mobil-
ity of opponent pieces around your king
We also used some well recognized heuristics for po-
sitional features. Some of these can be seen in Figure 3
and are listed below:

Figure 1: Attack, Defense, Mobility

After creating the aforementioned networks for each

board, we extract features from these network represen-
tations. Some of the features are:
• Number of edges in the Mobility network: indicator
of the mobility of various chess pieces on the board
(termed as ’activity’ in chess)
• Number of nodes in the Mobility network: captures
the notion of a ”space advantage” in chess Figure 3: Pawn structures
• Number of attack and defense edges: Each side will • Isolated pawns: Pawns that cannot be supported
ideally want to have a fewer attack edges from the by other pawns (indicates a weakness in the pawn
opponent and more support edges to its own pieces structure)
• Number of weakly connected components: This • Doubled pawns: Multiple pawns in the same col-
measures how well the pieces coordinate with each umn, again indicative of weaknesses in the structure
other as they barricade movement
 • Open files: A column on the chess board without After this, we retrained the features and compared per-
any pawns of the same color. Considered an advan- formance and repeated the feature selection process if
tage in presence of rooks, since it allows for addi- necessary. At last, we used the final model to evaluate
tional mobility the test error.
• Passed pawn: A pawn which cannot be attacked by
any enemy pawn and has cross the 5th rank on the
chess board. 5 RESULTS
In this section we discuss our prediction accuracy and
4 MODEL AND FEATURE SE- compare it with engine score as a predictor of the out-
come.
LECTION
In this section, we describe the models that we used for 5.1 Baseline Accuracy
our classification problem and the approach taken for For each board position we know what the engine-
feature selection. predicted score for that position is. Since a positive score
for one side indicates a stronger position for that side,
4.1 Models Used: we were able to measure a baseline prediction accuracy
using the engine’s scores as a proxy for the result of the
Considering the differences in discriminative and gener- game. In other words, if the engine had to predict the
ative models, we used the following classifiers for our result of a game by looking at the board positions, it
binary labelling task: would simply calculate the board score, s, and predict:
• Naive Bayes (Generative)

1 :s≥0
• Linear Discriminant Analysis (Generative) h(s) =
0 :s<0
• Logistic Regression (Discriminative)
• Support Vector Machines (Discriminative) This perceptron-like approach of calculating baseline
• Random Forests (Discriminative, Ensemble) accuracy allowed us to evaluate how well engine score
can predict game outcomes at different board scores.
The performance of the different models can be seen in
the Table 2 and have been discussed in greater detail in 5.2 Machine Learning Accuracy
the next section.
Consider the case when engine-predicted score is exactly
0 but the game ends decisively. Thus, engine predicts
4.2 Overview of Training, Testing and equal probability for both sides to win making the base-
Feature Selection: line accuracy as 50%. The prediction accuracies of our
method in this case are shown in Figure 2. Random
Forests performed the best with test accuracy of 63.06%
which is quite good as compared to baseline accuracy
50%.

Table 2: Machine Learning Accuracy Results

Classifier Train Test

1 Random Forest 62.91% 63.06%
Figure 4: Overview of approach to feature and model 2 Logistic Regression 56.99% 55.45%
selection 3 SVM (RBF) 56.19% 55.59%
4 LDA 57.97% 57.26%
Our first step involved mapping the data in FEN to 5 Naive Bayes 54.88% 53.88%
the set of features described earlier in addition to some
others by constructing the chess game networks. We Chess positions are often very complex to analyze us-
split this dataset into a training and cross-validation set ing this feature-based approach. Many features might
(70%), and a test set (30%). For each of the models indicate a weakness depending on a particular configu-
described above, we trained the model and optimized ration for the other features but a strength for another
hyper-parameters using the training and cross-validation configuration. For example, Open files may be a strong
dataset to prevent bias-variance issues. After analyz- advantage when a side has rooks on these files but can
ing the feature importance using data visualization tech- be a weakness if the side does not possess rooks but the
niques, we removed poor predictors and noisy features. opposition has rooks with the same set of open files. It is
very hard for most discriminative/generative models to
capture this complexity on a case by case basis, but de-
cision trees are intrinsically designed to handle such kind
of feature interactions. Thus, random forests is able to
outperform other classifiers.

5.3 Analysis of Results Figure 7: Augmenting Machine Learning with Current

Chess Engines
We considered board positions where the absolute value
of the engine score was less than some value x close to 0.
We varied the value of x from 0 to 1 and compared the 6 CONCLUSION & FUTURE
machine learning accuracy with the baseline accuracy for
that set of board positions. The results can be seen in
WORK
figure 5. When x is very close to 0, machine learning ap- Using machine learning we were able to predict the game
proach using positional features performs very well and outcome with approximately 63% accuracy for positions
outperforms the baseline accuracy with a good margin. where the engines were unable to assign an advantage to
When x is greater than 0.35, the baseline accuracy is either side (i.e. prediction accuracy of 50% if we use the
higher. This is because having positions with different engine score as a predictor). These results demonstrate
scores makes it harder for the features to capture the that engines lack important positional insight into the
variety of detail in the board positions. On the other game that our features and models capture. Given this,
hand, by training only on a small range of board scores possible future directions of our work are:
we are able to achieve much higher accuracies.
• Enhanced features: A case-by-case analysis of ex-
isting features with the possibility of additional fea-
tures would provide additional insight to incorpo-
rating positional play in chess engines.
• Augmenting chess engines with positional
play: The results from machine learning can be
added to the board evaluation function that chess
engines use and can be tested against state-of-the-
art engines. An example of how our predictions
can be used to recalculate board scores and im-
prove decision-making can be seen in Figure 7. In
this example, positions predicted to win and lose
are awarded and penalized 0.1 points respectively.
Note: Since feature computation can be expensive,
it should be included only at the top few levels in
the game-tree search rather than at the leaf-level.
Figure 5: Machine Learning vs Baseline
The source code for many state-of-the-art engines is
available to the public [3].
We also identified features that are most important
in predicting the board outcomes. The top five features
with their relative importance are shown in figure 6. We References
notice that mobility network features play an important
role in positional features. [1] H. van Kempen. (2006) Nunn tourna-
ments / cegt 40/120. [Online]. Avail-
able: http://www.husvankempen.de/nunn/
downloads/40 120 new/downloads.htm
[2] D. Farren, D. Templeton, and M. Wang, “Analysis of
networks in chess,” Stanford University, Tech. Rep.,
2013.

[3] [Online]. Available: https://github.com/mcostalba/

Stockfish
Figure 6: Relative importance of top features

View publication stats