You are on page 1of 22

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/275520452

EVALUATION OF CHESS GAMBITS

Research · April 2015

READS

119

1 author:

Jamal Munshi
Sonoma State University
54 PUBLICATIONS 50 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Jamal Munshi
letting you access and read them immediately. Retrieved on: 27 May 2016
EVALUATION OF CHESS GAMBITS

JAMAL MUNSHI

ABSTRACT: A dual engine experimental design previously described (Munshi, Evaluation of alternative continuations
of chess openings, 2015) is used to evaluate the effectiveness of chess gambits by comparing ten gambit
continuations with their closest mainline non-gambit twin in terms of the probability vector that generates chess
game outcomes. Six of the gambits were found to be failed innovations because they changed the probability vector
in favor of the opponent. The other four gambits were found to be benign innovations with two of them showing no
change in the probability vector and the other two showing that the probability vector was changed in a neutral
direction. More than these specific findings, however, the real purpose of the study is to offer an objective and
1
quantitative method that may be used to evaluate any gambit .

1. INTRODUCTION

The Queen’s Gambit is a famous opening sequence in chess in which White offers Black the c-pawn, an
offer that Black does well to decline (Ward, 2012) because its capture gives White a positional
advantage that exceeds the value of the sacrificial pawn. The success of the Queen’s Gambit offers chess
analysts the tantalizing possibility of other pawn sacrifices in the opening that can achieve the same kind
of positional advantage. A large number of such opening sequences, called “gambits” have been devised
and surely more of them will be devised in the future. Few gambits have enjoyed the success of the
Queen’s Gambit but a large variety of gambit innovations continue to be played at all levels of chess
(Shaw, 2013) (Eric, 2012) (Keene, 1993).

It is argued that gambits may be used to surprise opponents and force them into board positions with
which they are not familiar or to lure them into traps and early checkmates but no objective and rational
measure exists for evaluating the merit of gambits in terms of the probability of winning ceteris paribus.
It remains a contentious issue in chess whether the positional advantage gained by the pawn sacrifice
compensates for material loss2 and this issue is the essential research question of this paper.

A method for comparing alternative continuation lines of chess openings was described in a previous
paper (Munshi, Evaluation of alternative continuations of chess openings, 2015). In this study we show
that the methodology for comparing alternative continuation lines may be used to evaluate gambits
simply by describing the gambit line and its closest non-gambit continuation in the opening book as two
alternative opening line continuations to be compared. The method is demonstrated with ten gambits
selected for this study. A wide range of evaluations is observed in the selected gambits. The findings are
mostly consistent with expert opinion. More than the specific findings, however, the real value of this
work is the development of an objective methodology for the evaluation of chess gambits in general.

1
Date: April, 2015
Key words and phrases: Albin Counter gambit, Blackmar-Diemer gambit, Englund gambit, Benko gambit, Budapest gambit,
Danish gambit, King’s gambit, Elephant gambit, Latvian gambit, Evans gambit, chess gambits, chess openings, bootstrap, Monte
Carlo simulation, computational statistics, numerical methods, chess engines, engine analysis, chess analysis, trinomial process
Author affiliation: Sonoma State University, Rohnert Park, CA, 94928, munshi@sonoma.edu, ssrn.com/author=2220942
2
other than the Queen’s Gambit
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 2

2. THEORY

For the purpose of comparing opening lines, chess games may be viewed as a contest between the
white colored pieces and the black colored pieces and chess game outcomes may be thought of as a
trinomial stochastic process described by a three-dimensional probability vector π defined as

Equation 1 π = π(pw,pb,pd)

Here, pw is the probability that White will win, pb is the probability that Black will win, and pd is the
probability that the game will end in draw. This vector has two degrees of freedom because it is fully
specified by assigning values to any two of the three component probabilities by virtue of the constraint
that there is no other state possible at the end of a chess game. The three component probabilities must
therefore add up to one.

Equation 2 pw+pb+pd = 1

We assume that the probability vector π is defined by (1) the opening line played which we may refer to
as the opening effect (OE), (2) white’s first move advantage3 (FMA), (3) the overall move imperfection
rate in the game defined by the level of play (IMP), and (4) the difference in playing strength between
the player making the white moves and the player making the black moves (DIFF). We also assume that
there exist perfect and neutral openings such that OE=0, that is, the value of π is not changed by the
opening.

Suppose that there is no difference in playing strength between the two players (DIFF=0) and that the
move imperfection rate is zero (IMP=0). If a perfect and neutral opening line is played (OE=0)
π = [FMA, 0, (1-FMA)]. For example, if FMA=5% we can write π=[0.05,0,0.95]. Thus, in the perfect chess
game where the two players are exactly equally matched and neither makes a move error, and in which
the opening was perfect and neutral white holds a 5% probability of winning and the probability that the
game will end in draw is 95%. Black has no chance of winning under these conditions.

Now suppose that the players are not perfect and occasionally make mistakes but at a very small error
rate of 2%, that is there is a 2% probability that the game will be decisive by virtue of move errors. On
average, White will win half of these games and Black the other half. Under these conditions the value
of the probability vector is π=[0.06,0.01,0.93]. Black now stands a 1% probability of winning by virtue of
random move errors. White’s probability of winning has gone up to 6%. By subtraction, the probability
that the game will end in draw has dropped to 93%.

If the opening is not perfect and neutral then the value of π will be different by an amount that we may
ascribe to an opening effect (OE). Suppose that White selects a sub-optimal opening line that increases
Black’s probability of winning by 3 percentage points. This is the kind of opening effect or OE that our

3
Estimated to be about 4% to 7% incremental probability of winning (Munshi, Pairwise comparison of chess opening variations,
2014)
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 3

experiments are designed to detect. In all such cases the OE changes the probability vector because the
opening is not perfect and neutral. The value of the probability vector is now π=[0.06,0.04,0.90].

Even if we are unable to identify a perfect and neutral opening, we can still compare opening line
continuations. If the only difference between two experiments that generated the measured probability
vectors is the opening continuation line (CL) used, then differences between the two probability vectors
may be ascribed to the CL. For example if we find that π=[0.06,0.04,0.90] for CL#1 and
π=[0.16,0.04,0.80] for CL#2 we can conclude that CL#2 contains an advantage for white relative to CL#1.
It now remains for us to design engine experiments in which these equations may be applied.

3. METHODOLOGY

We use a dual engine design (DED) for our experiments with Komodo-8 (Kaufman, 2014)and Stockfish-5
(Romstad, 2014) as our two engines of choice. The DED decreases engine bias and generates a greater
variation in the moves played by the engines (Munshi, A test for engine bias, 2014). It has been shown
with engine experiments that these two engines are sufficiently close in playing strength under the
experimental conditions used so that the opening effect is not likely to be overwhelmed by difference in
playing strength if the openings are tested under identical conditions (Munshi, Evaluation of alternative
continuations of chess openings, 2015). These conditions are as follows.

1. The experiments are conducted by the Deep Shredder GUI (Meyer-Kahlen, 2014).
2. The engine parameters are set at the default values as shipped.
3. Eight threads are specified for both engines to match the number of physical CPU cores.
4. The hash table size is set to 7.68 gigabytes for both engines.
5. Time controls are set to 30 minutes for the first 30 moves and 10 minutes for the rest4.
6. There is no time increment per move. Each game is completed in 80 minutes or less.

Under these conditions the search depth reached by the engines in the first engine move after the
specified opening sequence had been exhausted was 25 to 30 plies for Stockfish and 22 to 26 plies for
Komodo. The search depths gradually increased as the game progressed. A sample size of 100 games
per opening continuation was selected for the experiments to strike a balance between statistical power
and the ability to complete the experiments in a reasonable length of time. It was also a consideration
that small differences in opening continuations detected with very high statistical power may not be
relevant to games played by humans. The move imperfection rate (IMP) in human games is likely to be
too high for small differences in opening effects to be important (Munshi, Pairwise comparison of chess
opening variations, 2014).

4
Many of the opening sequences tested are very short consisting of two to four half-moves and therefore it was deemed
necessary to front-load the time specification and give the engines more time to compute the opening continuation moves.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 4

Ten gambits are selected for testing – five from a queen’s pawn sequence and five from a king’s pawn
sequence. The selected gambits and the move sequences tested are shown in Table 1.

Gambit Gambit moves Gambiteer Non-gambit twin


5
Albin Counter Gambit 1. d4d5 2. c4e5 Black 1. d4d5 2. c4e6
6
Benko Gambit 1. d4Nf6 2. c4c5 (3. d5b5) Black 1. d4Nf6 2. c4e6
Blackmar Diemer Gambit 1. d4d5 2. e4 White 1. d4d5 2. c4
Budapest Gambit 1. d4Nf6 2. c4e5 Black 1. d4Nf6 2. c4e6
Englund Gambit 1. d4e5 Black 1. d4d5
Danish Gambit 1. e4e5 2. d4 White 1. e4e5 2. Nf3
Elephant Gambit 1. e4e5 2. Nf3d5 Black 1. e4e5 2. Nf3Nc6
Evans Gambit 1. e4e5 2. Nf3Nc6 3. Bc4Bc5 4. b4 White 1. e4e5 2. Nf3Nc6 3. Bc4Bc5 4. c3
King's Gambit 1. e4e5 2. f4 White 1. e4e5 2. Nf3
Latvian Gambit 1. e4e5 2. Nf3f5 Black 1. e4e5 2. Nf3Nc6
Table 1: The ten gambits selected for testing and their closest non-gambit twins

In six of the ten gambits, the gambiteer, that is the side that initiates the gambit, is Black. In the other
four gambits, the gambiteer is White. Each gambit line in Table 1 ends with the signature half-move that
defines the gambit. A special case is the Benko Gambit which is normally considered to be defined by 3.
d5b5 but it is defined differently for this study as 2…c5 which mirrors the non-gambit sequence 2…e6.
The move 2…c5 triggers the d5b5 sequence and in fact in all 100 games both engines played 3. d5b5. We
therefore take 2…c5 as an alternative to 2…e6 as the signature half-move of the Benko Gambit.

The gambit variations are compared with the opening book mainline continuations (Jones, 2014) shown
in the column labeled “Non-gambit twin”. The difference between the two sequences to be compared is
highlighted. Further continuation moves of the gambit lines are not specified because that would leave
open the possibility that the results are an artifact of the continuation specified rather than an
evaluation of the gambit itself. The reference non-gambit twin line is set to the same number of half
moves as the gambit line.

We compare the gambit and non-gambit lines as alternative continuations of an opening line (Munshi,
Evaluation of alternative continuations of chess openings, 2015). Each of the seventeen7 unique opening
sequences in Table 1 is tested in an engine match of 100 games between Komodo-8 and Stockfish-5
under controlled experimental conditions with each engine playing 50 games as white and 50 games as
black. The primary data recorded for each experiment are the number of wins by White and the number
of wins by Black. These two values serve as a two-dimensional measure of the effectiveness of the
continuation line being evaluated and are used to estimate the corresponding probability vector that
generates game outcomes under the specified conditions.

5
Both 2…e6 and 2…c6 are played but we have chosen 2…e6 as our baseline in this case.
6
The moves 3. d5b5 were not specified but were played by the engines in all 100 games.
7
Three of the non-gambit sequences are used twice.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 5

Each gambit line is compared with the corresponding non-gambit line in two dimensional Cartesian xy
coordinates with x = number of wins by White and y = number of wins by Black. The Euclidean distance
between the two continuation lines is used to determine whether they differ with respect to their effect
on the probability vector π. Both the magnitude and direction of the distance vector are important and
so the comparison is made in a two-step process. In the first step we determine whether the distance is
large enough; and if it is, the second step is necessary to interpret the distance in terms of its direction.

Designating d as the observed distance in the sample between a gambit line GL and its corresponding
non-gambit twin NGL, and δ as the true distance in the population8 between GL and NGL, we set up a
classical inferential hypothesis test for the magnitude of the distance δ. The null (H0) and alternate (Ha)
hypotheses for δ are stated as H0: δ=0 and Ha: δ>09.

We test the hypothesis at a maximum false positive error rate of α=0.001 as suggested by Valen Johnson
(Johnson, 2013). The standard deviation σ of the sampling distribution of d is estimated with a Monte
Carlo bootstrap procedure that creates one thousand simulated repetitions of the experiment. We then
compute t = d/σ and use the Student’s t-distribution to estimate p-value = the probability of observing a
distance ≥ d under conditions of H0. It is noted that ten comparisons are necessary in this study and that
the study-wide false positive error rate is likely to be much higher at approximately αS=0.01 according to
what is called the Bonferroni principle (Holm, 1979). The Bonferroni principle implies that there is a 1%
chance that at least one difference will be detected in random numbers if ten comparisons are made at
α=0.001.

In each comparison, if the p-value < α, we reject H0 and conclude that δ>0 and that therefore
π(GL)≠π(NGL), that is, the probability vectors that generate chess game outcomes for the gambit and
non-gambit lines are different. To understand the nature of this difference it is necessary to interpret
the direction of the distance vector according to Figure 1.

Figure 1: Direction map for interpreting δ>0

A distance vector from the non-gambit continuation NGL at the origin to GL located in the first quadrant
of the Cartesian coordinates shown in Figure 1 indicates that the π(GL) contains a higher value of pw and

8
Between the unobservable probability vectors that generated the sample data
9
The magnitude of the distance is an absolute value.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 6

also a higher value of pb and that therefore there is no difference between NGL and GL in terms of the
relative advantage of either color although it implies a higher probability of decisive games in the gambit
continuation. A similar neutral interpretation can be made if GL lies in the third quadrant relative to NGL
in which case pw and pb would both be lower for GL indicating a lower probability of decisive games for
the gambit line without a difference in relative advantage for either color. Quadrants 1 and 3 are
colored in yellow in Figure 1 to indicate that they are neutral directions in terms of relative advantage.

A very different situation arises if GL lies in the second or the fourth quadrant because these directions
imply that the ratio of pw/pb has changed to favor one color over the other. Figure 1 is arbitrarily color
coded from White’s perspective. The second quadrant is colored red because this direction implies a
reduction in White’s relative advantage. In this case, π(GL) contains a higher value of pb and/or a lower
value of pw so that GL offers a relative advantage to black when compared with NGL. The opposite is
true in the 4th quadrant and it is colored green because it offers a relative advantage to White.

To determine which quadrant represents the direction of δ we look at the 95% confidence interval of
the angle subtended by the distance vector. If the interval lies wholly in one quadrant then a decision
based on direction is possible but if the interval straddles two different quadrants then no quadrant can
be inferred and no decision based on direction can be made. For example if the direction interval is 95o
to 165o we can identify the direction as the 2nd quadrant and a decision based on direction is possible.
However, if the interval is 165o to 200o the direction interval straddles the 2nd and 3rd quadrants. In this
case, the direction cannot be identified and the test is therefore inconclusive.

It is noted that although our α-level has been set to α=0.001 for the magnitude hypothesis, the
confidence interval for direction is set to α=0.05. This change was made because at lower values of α,
the direction intervals are mostly greater than 90o which makes it impossible to identify the direction in
terms of a single quadrant. The test for magnitude is our primary criterion for the detection of
differences in π. The angle is a subordinate measure used only after a difference in π is detected. It is
used to estimate a general direction of the distance vector. There was only one instance in which the
uncertainty in direction was a factor in the evaluation of a gambit10. The standard deviation of the
direction is computed with the same Monte Carlo bootstrap procedure used for estimating the standard
deviation of distance. The computational details are included in the data archive for this paper (Munshi,
Gambit paper data archive, 2015).

If the p-value > α, we fail to reject H0. In this case the evidence does not show that π(GL)≠π(NGL) and the
direction of the distance vector has no rational interpretation as we must allow for the possibility that
δ=0 and that π(GL)=π(NGL). The data do not provide sufficient evidence that chess game outcomes for
the gambit and non-gambit lines are not driven by the same underlying and unobservable probability
vector. If the p-value is sufficiently high, we may use a form of Bayesian logic to accept H0 and conclude
that π(GL)=π(NGL). For this study we have somewhat arbitrarily set that probability to β=100*α, or
β=0.10. When using both α and β, our complete hypothesis proceeds as follows: if p-value < α, reject H0

10
The Evans Gambit
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 7

and conclude that π(GL)≠π(NGL) and if p-value > β accept H0 and conclude that π(GL)=π(NGL); otherwise
fail to reject H0 in which case the test is inconclusive.

Additional data are taken to assess the variability in the opening moves computed by the engines after
the specified moves in Table 1 are exhausted. The opening variability (OV) of the engines is measured by
counting the number of unique moves played by the engines from the end of the moves specified in
Table 1 up to the 10th move by Black in the 100 games played in each experiment. The opening phase of
the game is assumed to consist of 20 half-moves for this analysis.

As an example, three half-moves are specified in Table 1 for the Albin Counter Gambit leaving 17 half
moves to be computed by engines in the first 20 half moves of the game. In this case, a value of OV=17
would indicate that there were no variations played and the very same 17 half-moves were played in all
100 games. By contrast, a value of OV=500 would indicate that a great variety of moves were played by
the engines in the opening 20 half-moves of the game. High values of OV serve as a subjective measure
of the extent to which the experimental results were not influenced by engine bias. In addition to a
numerical value of OV for each continuation tested we also present the actual opening moves played by
the engines in an opening book format so that experts and analysts who have studied these gambits can
more easily check the engine moves against their own analysis.

4. DATA ANALYSIS

The results of the 17 engine experiments are summarized in Table 2. The data archive for this paper
contains all 1700 games in PGN format with the first 20 half-moves of each experiment also presented in
BKT format (Munshi, Gambit paper data archive, 2015).

Gambit games won NonGambit games won


Gambit Gambiteer White Black OV White Black OV
Albin Counter Gambit Black 45 3 398 19 10 1253
Benko Gambit Black 34 12 969 17 17 1490
Blackmar Diemer Gambit White 11 31 970 21 16 1320
Budapest Gambit Black 42 6 317 17 17 1490
Englund Gambit Black 83 2 470 16 22 1293
Danish Gambit White 14 16 635 25 26 1267
Elephant Gambit Black 53 3 264 25 26 1075
Evans Gambit White 24 10 546 27 25 811
King's Gambit White 27 28 695 25 26 1267
Latvian Gambit Black 72 6 188 25 26 1075
11
Table 2: Results of 17 engine experiments with n=100 games per experiment

We compute the Euclidean distance between each gambit and its non-gambit twin and use a Monte
Carlo procedure to generate a simulated sampling distribution of Euclidean distances to determine

11
The 20 results tabulated represent 17 experiments. Some non-Gambit results are used more than once.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 8

whether the observed distance could have occurred by virtue of sampling variation even if H0 is true and
δ=0.

For example, in the case of the Albin Counter Gambit , the Euclidean distance observed in our sample of
100 games is d=√((45-19)2+(3-10)2) = 26.93 games which may be represented as 26.93% of sample size.
However, if we repeated these two experiments, it is unlikely that we would get exactly the same results
and the same distance.

In fact, by using the sample data as our best unbiased estimate of π, we can run simulated repetitions of
the experiments to see how different the results and the distances could be. Figure 2 is a screenshot of
such a simulation. It shows for example that a repetition of the non-gambit line could yield 25 wins by
white and 9 by black. The variation among the repetitions is a function of sample size. Smaller sample
sizes yield greater variation and larger sample sizes may be used to control this variation and gain
greater statistical power. The variation shown in Figure 2 derives from our sample size of n=100 games
per experiment. In the few simulated repetitions in Figure 2 we see that the observed distance could
vary from 18% to 34%. For this study, we generate 1000 simulated repetitions and use the standard
deviation of the 1000 distances as a proxy for the standard deviation of the sampling distribution of
distances. The Microsoft Excel spreadsheets used to create these simulations may be downloaded from
the data archive for this paper (Munshi, Gambit paper data archive, 2015). There are two spreadsheets
in the data archive, one for the 1. d4 openings and one for the 1. e4 openings.

Figure 2: Screenshot showing simulated repetitions of the experiments

Figure 3 through Figure 12 contain graphical depictions of the simulated replications of each pair of
experiments. The gambit continuation is shown in red and the non-gambit continuation is in blue. The
simulated sampling distribution of distances between them is used to test for a difference in π with the
null hypothesis H0: π(Gambit)=π(non-Gambit). The testable implication of this hypothesis is that δ=0.

If we find that π(Gambit)≠π(non-Gambit)then we classify the gambit as A=successful gambit if the


difference in π favors the gambiteer and as F=failed innovation if the difference in π favors the other
color. If the difference in π does not favor either color or if we do not find that that π(Gambit)≠π(non-
Gambit), then we classify the gambit as C=benign innovation, that is we have no evidence that the
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 9

gambit innovation changes the probability that the gambiteer will win ceteris paribus. All data and
computational details for all ten hypothesis tests presented in Figures 3 through 12 are available for
download in the data archive for this paper (Munshi, Gambit paper data archive, 2015).

4.1 Albin Counter Gambit: Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The 95% confidence interval for the angle of the Gambit is 271o to 300o. The Gambit variation lies in the
4th quadrant implying that the Gambit offers a relative advantage to White. Since the gambiteer is
Black, the Albin Counter Gambit is classified as F=failed innovation.

Expert opinions
Despite its apparent weakness, the Albin Counter Gambit may offer an attractive option for Black to test
the opponent’s skills in what may be an unfamiliar board position for many chess players
(thechesswebsite, 2009) (Martin, 2014). A specific strength of the Albin Counter Gambit is that it
threatens White with the Lasker Trap (ChessNetwork, 2012).

Lines played by the engines


In all 100 games of the Albin Counter Gambit experiment, the engines played 3. dxe5 followed by 3... d4
or Ne7. In the 3…d4 line 4. Nf3Nc6 was played. In the 3…Ne7 line the Queens were exchanged off with
4. cxd5 Qxd5 5. Qxd5 Nxd5.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 10

4.2 Benko Gambit Classification = C

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value > α fail to reject H0 Since p-value < β fail t accept H0
Decision: The test is inconclusive.
The data do not provide sufficient evidence that (Gambit) ≠ π(non-Gambit)
The angle of the Gambit is irrelevant in this case. The evidence does not show that the gambit has
changed the probability vector π. We therefore classify the Benko Gambit as C=benign innovation.

Expert opinions
Our failure to distinguish between the Benko Gambit and the mainline 2…e6 is consistent with the
opinion of some analysts who have identified specific strengths in the Benko Gambit line (Alburt, 2012).

Lines played by the engines


In all 100 games of the Benko Gambit experiment, the engines played 3. d5 b5 followed by 4. Bf4, Nf3,
Qc2, or cxb5.

4.3 Blackmar-Diemer Gambit Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 11

Check direction of distance vector for effect of δ>0


The 95% confidence interval for the angle of the Gambit is 116o to 176o. The Gambit variation lies in the
2nd quadrant implying that the Gambit offers a relative advantage to Black. Since the gambiteer is
White, the Blackmar-Diemer Gambit is classified as F=failed innovation.

Expert opinions
Some analysts promote the Blackmar-Diemer Gambit as a more aggressive game than the mainline
Queen’s Gambit (thechesswebsite, 2010), while others are less enthusiastic (Dzindzichashvili, 2014).

Lines played by the engines


In all 100 games of the Blackmar-Diemer Gambit experiment, the engines played 2... dxe4 followed by
3. Nc3Nf6 or 3. Ne2e6 4. Nbc3f5.
.
4.4 Budapest Gambit Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The 95% confidence interval for the angle of the Gambit is 275o to 313o. The Gambit variation lies in the
4th quadrant implying that the Gambit offers a relative advantage to White. Since the gambiteer is
Black, the Budapest Gambit is classified as F=failed innovation.

Expert opinions
An advantage of the Budapest Gambit is that if the White player is not familiar with this opening there is
a possibility of an early checkmate by Black (thechesswebsite, 2009). Some analysts recommend this
gambit for players at the amateur or club level of play because of the number of ways in which White
can go wrong in this opening (Seirawan, 2013).
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 12

Lines played by the engines


In all 100 games of the Budapest Gambit experiment, the engines played 3. dxe5 Ng4 followed by 4. Bf4,
e3, Nf3, Nc3, or Bd2. All 100 games in the engine experiment are available for download in PGN format;
and the first ten moves played are also available in opening book format in the data archive for this
paper (Munshi, Gambit paper data archive, 2015).

4.5 Englund Gambit Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The 95% confidence interval for the angle of the Gambit is 280o to 293o. The Gambit variation lies in the
4th quadrant implying that the gambit offers a relative advantage to White. Since the gambiteer is Black,
the Englund Gambit is classified as F=failed innovation.

Expert opinions
Despite its recognized weakness, there are some lines in this opening that can trap White players who
have not studied the Englund Gambit and that may be sufficient reason for Black to try this opening at
the club level of play (Jeffreys, 2007).

Lines played by the engines


In all 100 games of the Englund Gambit experiment, the engines played 2. dxe5 followed by 2…Nc6,
2…d6, or 2…a6.
.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 13

4.6 Danish Gambit: Classification = C

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The direction interval of the Gambit, between 180o and 264o, lies entirely in the neutral 3rd quadrant. The
effect of the gambit is therefore simply to lower the probability of decisive games without a change in
relative advantage. The Danish gambit is therefore classified as C=benign innovation.

Expert opinions
It is claimed that in the Danish gambit White has little to lose and much to gain at the club level of play
in terms of trap opportunities (ivanovitchs, 2009) (JonesMaster, 2009). The Danish is also considered by
many to be an important part of the opening repertoire for all players because of the inherent strengths
of certain lines of the opening for White in terms of rapid piece deployment and early attack
opportunities (jrobichess, 2007) (Matojelic, 2011) (Lutes, 1989) (Butler W. , 1966) (Goldner, 2015). On
the other hand there are some who feel that after 1. e4e5, black can gain an advantage if White does
not follow the Ruy Lopez line (Bologan, 2014).

Lines played by the engines


In all 100 games of the Danish Gambit experiment, the engines played 2… exd4 followed by 3. c3,Nf3, or
Ne2. After 3. c3 dxc3, 4. Nxc3 was played in all 100 games. The interesting sequence 3. c3 dxc3 4. Bc4
cxb2 Bxb2, of which so much has been written, was not played by the engines.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 14

4.7 Elephant Gambit: Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The direction of the Gambit, between 295o and 326o, lies squarely within the 4th quadrant which
indicates a relative advantage for White. Since the gambiteer is Black, the Elephant Gambit has failed
and is therefore classified as F = failed innovation.

Expert opinions
Expert opinion is mostly on the downside for the Elephant Gambit with the most positive aspect claimed
being that at the club level of play it could be used to surprise opponents who are not familiar with the
book lines for this opening (Flear, 2010) (Harding, 1997) (Kenilworth Chess Club, 2012). It is generally
recognized that the Elephant Gambit can lead to some very entertaining and unusual move sequences
(MatoJelic, 2011) (Seirawan, Elephant Gambit, 2014) and that for players who have not studied this
opening, there are a number of ways for White to go wrong and hand Black an early checkmate
(MatoJelic, 2011b).

Lines played by the engines


In all 100 games of the Elephant Gambit experiment, the engines played either 3. Nxe5 or 3. exd5. The 3.
Nxe5 line continued with 3…Bd6, 4. d4 dxe4 5. Bc4 Bxe5 6. Qh5 Qe7 7. Qxe5 Qxe5 8. dxe5 and the
queens were exchanged off. In the 3. exd5 line 3…e4 4. Qe2 Nf6 or 3…Nf6 4. Nxe5 Qxd5 followed by 5.
d4 or Nf3 were played.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 15

4.8 Evans Gambit: Classification = A

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)

Check direction of distance vector for effect of δ>0


The direction interval of the Gambit between 215o and 282o straddles the 3rd and 4th quadrants. The test
is inconclusive because the direction of the distance vector is unclear. It is classified as C = benign gambit
because the data do not indicate a strength or a weakness for the gambiteer.

Expert opinions
Expert opinion is universally positive for the Evans Gambit with many analysts suggesting that it is a
stronger line than the conventional Italian game (Moody, 1995) (Schiller, 2013) (Rohde, 1997) (Kelley,
2012) (Butler, 2009).

Lines played by the engines


In all 100 games of the Evans Gambit experiment, the engines played 4. Bxb4 followed by 5. c3 and
5…Ba5, Bd6, or Be7.

4.9 King’s Gambit: Classification = C


EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 16

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value > α, fail to reject H0 Since p-value > β, accept H0
Decision: δ=0 Conclude: π(Gambit)=π(non-Gambit)

Check direction of distance vector for effect of δ>0


The direction of the distance vector is irrelevant in this case.

Expert opinions
There are strong opinions on this gambit on both sides of the issue. The King’s Gambit was refuted by
Bobby Fischer in his article “A Bust to the King’s Gambit” (Fischer, 1961). More recently chess engine
programmer and International Master Vasik Rajlich used an extensive computer based study to support
Fischer’s position (Chessbase, 2012) and Grandmaster Eugene Perelshteyn, a recognized expert and
frequent commentator on chess openings, has described specific weaknesses in the King’s Gambit
(Perelshteyn, 2015). Yet, the gambit is seen at all levels of play with many analysts viewing the gambit as
an aggressive game with great diversity of playable lines many of which offer specific advantages to
White particularly if the opponent has not studied ways to defend against them (thechesswebsite, 2009)
(Kelley, King's Gambit, 2012) (Gallagher, 1993) (Johansson, 2005) (Hooper, 1996)12.

Lines played by the engines


In all 100 games of the King’s Gambit experiment, the engines played 2…exf4. White responded with 3.
Nf3 in 90 games. The usual King’s gambit sequence 3. Nf3g5 is seen in 59 games. The 3. Bc4Qh4
sequence made famous by Bobby Fischer was played in the remaining 10 games.
.
4.10 Latvian Gambit: Classification = F

Test for magnitude of the Euclidean distance


H0: δ=0 Ha: δ>0 α=0.001 β=0.10
Since p-value < α, reject H0 Decision: δ>0 Conclude: π(Gambit) ≠ π(non-Gambit)
12
A previous study in this series in which a specific line of the gambit was specified to six half moves had reported a weakness
in the King’s’ Gambit with the engine Houdini playing both sides of the board (Munshi, A method for comparing chess openings,
2014). The difference may be due to a combination of engine bias and a weakness in the line chosen. This issue is explored in a
forthcoming paper.
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 17

Check direction of distance vector for effect of δ>0


The direction interval of the Gambit between 283o and 302o lies in the 4th quadrant. The direction
indicates a relative advantage for White. Since the gambiteer is Black, the gambit has failed. The Latvian
gambit is thus classified as F = failed innovation.

Expert opinions
Analyst opinions on the Latvian Gambit seem to be that the value of the gambit lies in the its rarity and
that therefore at less than the Master level of play it can place White in unfamiliar territory (Butler K. ,
Latvian Gambit, 2014). For example, in 1e4e5 2Nf3f5 3Nxe5Bc5 4exf5 Black gains a dominant position by
making yet another sacrifice (Sturdivant, 2014). Yet, it is precisely in this argument that the fundamental
weakness of the gambit is revealed because the supposed advantage for Black requires White’s
cooperation. However, it is not unusual to find highly ranked players who are attracted to the sacrificing
style of play used by Black in this opening (Kosten, 2003) (Silman, 1998).

Lines played by the engines


After 1. e4e5 2. Nf3f5, White played 3. Nxe5 and black replied with either 3…Qf6 or 3…Nf6. The move
3…Bc5 was not played.

5. SUMMARY AND CONCLUSIONS

The results for all ten gambits are summarized in Table 3 sorted by Euclidean distance. Six of the ten
gambits are found to be failed innovations because the gambit innovation when compared with the
corresponding non-gambit continuation cedes an unassailable advantage to the opponent. The
Studentized distance shown in the table serves as a measure of the degree of the disadvantage to the
gambiteer in these failed innovations. In all of these cases, the direction of the distance vector lies in the
2nd or 4th quadrant indicating a difference in relative advantage between the non-gambit continuation
and the gambit continuation.

In two the remaining four gambits tested the observed Euclidean distance between the gambit
continuation and the reference non-gambit continuation was small enough to be explained by sampling
variation. In these cases, the King’s Gambit and the Benko Gambit, the data did not provide sufficient
evidence that the population distance δ > 0 or that the two experimental results being compared were
not generated by the same underlying probability vector. The observed Euclidean distance was too small
to reject the null hypothesis that π(Gambit)=π(nonGambit) and in the case of the King’s Gambit it was
small enough to accept the hypothesis that π(Gambit)=π(nonGambit).

The case of the Danish Gambit and the Evans Gambit is very different. Here we found sufficient evidence
that δ > 0 and that therefore π(Gambit)≠π(nonGambit) but the direction of the distance vector could not
be interpreted in terms of relative advantage. In the case of the Danish Gambit the direction was found
to be in a neutral direction that implied that the gambit continuation differs from the mainline non-
gambit continuation only with respect to the probability of decisive games without a relative advantage
to either color. The results for the Evans Gambit are unclear and inconclusive. We know that the Evans
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 18

gambit changes the probability vector but the uncertainty in the direction prevented a conclusion with
respect to relative advantage. Further research is planned to lower the degree of uncertainty in
direction for this gambit.

More than these specific conclusions, however, the primary objective of this work is to offer a
methodology that all chess players may use to evaluate gambits. The long run objective of this line of
research is to improve the quality of information available to chess players and at the same time help to
refine the opening book.

Table 3: Summary of results

6. REFERENCES

Alburt, L. (2012). The Benko Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=53QVuEvNETY

Bologan, V. (2014). Win if White avoids the Ruy Lopez. New in chess.

Butler, K. (2009). Evans Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=HuHkMom9oy0

Butler, K. (2014). Latvian Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=R5oI4gupzBU

Butler, W. (1966). A Danish Gambit. Peter Owen.

Chessbase. (2012). Busting the King's Gambit. Retrieved 2015, from Chessbase:
http://en.chessbase.com/post/rajlich-busting-the-king-s-gambit-this-time-for-sure

ChessNetwork. (2012). Lasker Trap. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=s_ED8vrQQlM

Dzindzichashvili, R. (2014). Blackmar Diemer Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=93LbTL7Y15w
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 19

Eric, S. (2012). The Blackmar-Diemer Gambit. Ishi Press.

Fischer, R. (1961). A bust to the King's Gambit. Retrieved 2015, from academicchess.org:
http://www.academicchess.org/images/pdf/chessgames/fischerbust.pdf

Flear, G. (2010). Starting out: Open games. Everyman Chess.

Gallagher, J. (1993). Winning with the Kings Gambit. Henry and Holt.

Goldner, M. (2015). Danish Gambit Accepted. Kindle.

Harding, T. (1997). Going on an elephant hunt. Retrieved 2015, from chesscafe:


http://www.chesscafe.com/text/kibitz15.txt

Holm, S. (1979). A simple sequentially rejective multiple test procedure. Holm, S. (1979). "A simple
sequentially rejective multipScandinavian Journal of Statistics, 6 (2): 65–70.

Hooper, D. (1996). The Oxford companion to chess. Oxford: Oxford University Press.

ivanovitchs. (2009). Danish Gambit Trap. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=Bf3hfOFfDao

Jeffreys, M. (2007). Traps in the Englund Gambit. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=5FO2cIlD9Rs

Johansson, T. (2005). The fascinating King's Gambit. Trafford Publishing.

Jones, R. a. (2014). Game database. Retrieved 2014, from chesstempo.com:


http://chesstempo.com/game-database.html

JonesMaster. (2009). Danish Gambit Trap. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=6DmNiGqgMcY

jrobichess. (2007). Danish Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=mCP9hgqOk6E

Kaufman, L. (2014). Komodo Chess. Retrieved 2014, from komodochess.com: http://komodochess.com/

Keene, R. (1993). The complete book of gambits. Henry Holt.

Kelley, D. (2012). Italian game: Evans gambit. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=J39v_OHeb3c

Kelley, D. (2012). King's Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=kViwcjLy1eQ

Kenilworth Chess Club. (2012). Elephant Gambit. Retrieved 2015, from Kenilworthian:
http://kenilworthian.blogspot.com/2012/12/elephant-gambit-c40-bibliography_12.html
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 20

Kosten, T. (2003). The Latvian Gambit lives. Batsford.

Lutes, J. (1989). Danish Gambit. Chess Enterprise.

Martin, A. (2014). Tremendous pressure with the Albin Gambit. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=8tZ65TNTwt0

Matojelic. (2011). Danish Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=JAI5r2jSx5M

MatoJelic. (2011). Elephant Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=oe9ltv1L6F4

MatoJelic. (2011b). Elephant Gambit: David Bronstein. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=ZZcnB1bLky4

Moody, R. (1995). The Evans gambit revolution. Chess Digest.

Munshi, J. (2014). A method for comparing chess openings. Retrieved 2015, from ssrn.com:
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2415203

Munshi, J. (2014). A method for comparing chess openings. Retrieved from ssrn.com:
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2415203

Munshi, J. (2014). Comparing chess openings part 3. Retrieved 2015, from ssrn.com:
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2441568

Munshi, J. (2014). Pairwise comparison of chess openings. Retrieved 2014, from ssrn.com:
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2472783

Munshi, J. (2015). Evaluation of alternative continuations of chess openings. Retrieved 2015, from
ssrn.com: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2553614

Munshi, J. (2015). Gambit paper data archive. Retrieved 2015, from Dropbox:
https://www.dropbox.com/sh/n88puzd2licg2f2/AACzdS34EELAq4HMhE8M2Lpxa?dl=0

Perelshteyn, E. (2015). Beat the King's Gambit. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=KZt2eaW_K24

Rohde, M. (1997). The great Evans gambit debate. Thinkers Press.

Romstad, T. e. (2014). Stockfish. Retrieved 2014, from Stckfish: https://stockfishchess.org/

Schiller, E. (2013). Chess superstars play the Evans gambit. Create Space.

Seirawan, Y. (2013). Exploring d4 Budapest Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=_SvgVk_6FNY
EVALUATION OF CHESS GAMBITS, JAMAL MUNSHI, 2015 21

Seirawan, Y. (2014). Elephant Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=7hJGuzguxvQ

Shaw, J. (2013). The King's Gambit. Quality Chess.

Silman, J. (1998). The complete book of chess strategy. Siles.

Sturdivant, G. (2014). Latvian Gambit trap. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=-viCr8TnKmA

thechesswebsite. (2009). Albin Counter Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=VDfjsMXlcek

thechesswebsite. (2009). Budapest Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=urITvqQleUE

thechesswebsite. (2009). Chess openings - the King's Gambit. Retrieved 2015, from Youtube:
https://www.youtube.com/watch?v=n5ImJ2row6E

thechesswebsite. (2010). Blackmar Diemer Gambit. Retrieved 2015, from Youtube:


https://www.youtube.com/watch?v=BakZpcj_Yek

Ward, C. (2012). Play the Queen's Gambit. Everyman Chess.

You might also like