Backgammon: The Optimal Strategy

for the Pure Running Game∗

Edward O. Thorp

[Edward O. Thorp & Associates]

Abstract. The pure running game (end game, no-contact game) occurs in backgam-
mon when no further contact between opposing men is possible. It compares roughly
in importance to the end game in chess. This paper presents a complete exact so-
lution to the pure running game with the doubling cube under the simplifying as-
sumption (Model 1! ) that dice totals can be arbitrarily subdivided in moving men
and that men can be borne off even though not all are in the inner table.
The actual pure running game (i.e., without these simplifying assumptions) can
be solved to high approximation by the methods of this paper. The strategies ob-
tained are superior to those employed by the world’s best players.
Our strategies and methods for the end game lead to substantially improved
middle-game (where contact is still possible) play. It is feasible to find exact solutions
to some middle-game situations. By also employing positional evaluation functions,
it seems feasible to program an existing computer to become the world backgammon
champion.

1 Introduction

The game of backgammon, along with chess and checkers, is among the old-
est games known. A primitive backgammon-like game dated 2600 B.C. is on
display in the British Museum (Jacoby and Crawford 1970, pp. 8–9).
In place of the traditional board we use the equivalent representation of 26
linearly ordered cells numbered 0, 1, . . . , 25. The traditional board corresponds
to cells 1, . . . , 24. White men travel from higher to lower numbered cells on
White turns and Black men move in the opposite direction. To preserve the
symmetry, we also use the coordinate systems i = W i = B(25 − i). See Figure
1.
The object of White is to move all men to W 0 and the object of Black is
to move all men to B0. The first player to get all men to his 0 cell wins. The

∗ Presented at the Second Annual Conference on Gambling, South Lake Tahoe, Nevada,
June 15–18, 1975.
 238

White "bears off" here; these W Traditional board White "bar"

men never move again
ends here
White’s home White’s outer
or inner board board
W

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Edward O. Thorp

25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
B

Black’s outer Black’s home

Black "Bar" Traditional board board or inner board
(Black blots
ends here Black "bears off" here
sent here when hit)
Traditional board "folded" here

Fig. 1. Backgammon board. White travels left on his turn. Black travels right on his turn.
 Backgammon: The Optimal Strategy for the Pure Running Game 239

men are initially as in Figure 1. The additional rules (see, e.g., Holland 1973,
Hopper 1972, Jacoby and Crawford 1970, Lawrence 1973, Obolensky and
James 1969, or Stern 1973) are assumed known to the reader and will be
used, without a complete formal statement, as we proceed.
[In Section 2 we give bounds for the duration of simplified models of the
game. In Section 3 we give bounds for the number of board configurations.
Section 4 gives a recursive solution to the Model 1 end game, i.e., the one-
checker pure race model, and Section 5 extends this to Model 1! , which adds
the doubling cube. Section 6 discusses possible extensions of the research.]

2 Duration of the game

It is clear from Figure 1 that each player requires a minimum of 167 steps to
bear off all his men. The expected number of steps on the first player’s first
turn is 7; for each subsequent roll of the dice pair it is 8 1/6. Dividing 167 by
8 1/6 gives 20.45 so it is plausible to suppose that the lower bound for the
expected duration of the game might have order of magnitude 20.
More precisely, let X1 , X2 , . . . , Xn , . . . be the totals rolled by the second
player. The Xi are independent identically distributed random variables. Let
M = 167 and let n∗ be the least n such that X1 + · · · + Xn ≥ M . Then E(n∗ )
is the expected time for player 2 to finish assuming no “interference” and no
“waste” (i.e., Model 1, below). Thus E(n∗ ) ≥ 20.45 in this case and à fortiori in
the standard game. Similarly for the first player (E(m∗ ) − 1)E(Xi ) + 7 ≥ M
or E(m∗ ) ≥ 20.59. [Define the duration of the game to be min(m∗ , n∗ ).]
Note, however, that the expected duration of the standard game, or even of
Model 1 below, could still be less than 20.45 because the game ends when
either player finishes and in general E(min(m∗ , n∗ )) < min(E(m∗ ), E(n∗ ))
for random variables m∗ and n∗ .
Definition 1 (Model 1). The random walk idealization of the game assumes
that two players (F and S for “first” and “second”) alternately roll the pair
of dice. They start with totals of zero and accumulate points according to the
outcomes of their rolls, valued as is done subsequent to the first roll of the
actual game. The goal of F is a total N (F ); that of S is N (S). The first
player to accumulate a total which is greater than or equal to his goal wins.
Model 1 omits several features of the game: (1) there is no blocking and
no hitting of blots; in Model 1 opposing men can both occupy the same cell;
the “flows” of opposing men do not interact; (2) there are no restrictions on
bearing off in Model 1 whereas in the game all men not already borne off must
be in the player’s home board before he can continue to bear off; (3) Model 1
allows the total to be arbitrarily subdivided for use in moving men whereas in
the game each move of a man must be in a jump that uses the entire number
on a die. For instance, rolling (5, 3) the admissible jumps are 5 and 3. Rolling
(4, 4) (“doubles”) they are 4, 4, 4, and 4.
240 Edward O. Thorp

Model 1 may be played as a race between F and S, each on his own

individual track. The tracks have N (F ) + 1 and N (S) + 1 cells, respectively.
The initial configuration is a single man in cell 0 of each track. Players roll
alternately, moving the total number of steps indicated by the dice. The first
player to reach the last cell of his track wins.
The expected duration E(D1 ) of the Model 1 game is a lower bound for
the expected duration E(DG (s)) of the standard game G where s is any pair
of strategies for Black and White. To see this, note that in Model 1 each
roll of the dice is fully utilized in reducing the corresponding player’s pip
count (pips needed to win) whereas in the standard game part or all of the
roll of the dice may not be utilized. Reasons include blocked points and the
possible waste of some of the roll in bearing off. Also the player on roll can
hit the other player’s blots thereby increasing the other player’s pip count.
Thus each Model 1 player’s pip count will always be less than or equal to the
corresponding player’s standard game pip count so the Model 1 game will end
no later than the standard game. So D1 ≤ DG (s) for each realization of the
standard game hence E(D1 ) ≤ E(DG (s)) for all pairs of player strategies s.
Model 1 corresponds to a random path on the lattice points of the non-
negative quadrant with absorption at x = 167 (F wins) and at y = 167
(S wins). The odd numbered steps are horizontal and to the right distances
X1 , X3 , . . ., and the even numbered steps are vertically up distances Y2 , Y4 , . . .,
where X1 , X3 , . . . and Y2 , Y4 , . . . are independent random variables distributed
like the dice outcomes. A computer simulation will now readily yield a fairly
accurate value of E(D1 ).
What is the least number n of moves possible in the standard game? We
know n ≥ $167/24% = 7 in the Model 1 game because 24 points (double sixes)
is the maximum per roll. (The symbol $x%, the “ceiling” of x, is the smallest
integer k such that k ≥ x.) Hence 7 is a lower bound.
To see that the game cannot end in 7 moves, consider a somewhat more
realistic model.
Definition 2 (Model 2). The no-contact model assumes the usual valuation
of, and use of the outcomes of the rolls. Hence the men must move in jumps
corresponding to the dice faces. This is the only difference from Model 1. A
board configuration and who has the move is specified. There are no stakes
and no doubling cube. Men can be borne off even though not all have reached
the inner table. Unlike Model 1, there may be “waste” when men are borne
off: If a six is used to bear off a man from cell 5, the unused step cannot be
applied to another man.
Visualize Model 2 as a game where the players’ pieces are on two separate
boards. Men then move as in the standard game except that not all men
need to reach the inner table before men can be borne off. This can be an
advantage. For instance suppose W has only two men, on W 1 and W 13.
Now W rolls (6, 1). He has a choice of W 1 → W 0 and W 13 → W 7 or of
W 13 → W 6 whereas in the standard game only the latter is allowed. But
 Backgammon: The Optimal Strategy for the Pure Running Game 241

the configuration “one man on W 7” is better than “men on W 1, W 6” in that

n∗ (W 7) ≤ n∗ (W 1, W 6) and strict inequality holds with positive probability.
This will be proven (easily) later. For now, note that a roll of (5, 2) bears off
one man on W 7 but with men on W 1 and W 6 it leaves a man on W 1.
In the preceding example, the Model 2 player upon rolling (6, 1) had a
choice as to whether to change {W 1, W 13} into {W 0, W 7} or into {W 1, W 6}
and we saw that {W 0, W 7} was better. The game situations in which a player
has a decision to make are “states.” The rule the player uses to make his
choices in each state is his strategy.

Definition 3. A configuration is an arrangement of the 15 White men and

the 15 Black men on the extended board of 26 ordered cells. A position is a
configuration and the specification of who has the doubling cube and what its
value is. We always assume the initial stake is 1 unit. A state is a position
and the specification of who is to move.

[A configuration may be thought of as a photograph of the board without

the doubling cube. Then a position is a photograph that also shows (the upper
face of) the doubling cube.]
The game can be unambiguously continued when the state is specified.

Definition 4. A strategy for a player is a rule that specifies how the player
is to move for each state of the game in which he is to move. For a given
state the strategy may be pure or mixed. In a pure strategy the player always
makes the same moves for a given roll. In a mixed strategy he chooses his
moves for a given roll from among several alternatives, each having a specified
probability of being chosen. Strategies also may be stationary or nonstationary.
A stationary strategy is one which does not depend on the time (i.e., turn
number). Each time a state repeats, a stationary strategy has the same rule
for selecting moves. A stationary pure strategy selects exactly the same moves.
A nonstationary strategy does depend on the time. A strategy pair s = (s1 , s2 )
is a specification of strategies for players 1 and 2 respectively.

We shall generally limit ourselves to stationary pure strategies in subse-

quent sections. This is justified by the facts that (1) there is a strategy pair
s = (s∗1 , s∗2 ) such that s∗1 and s∗2 are optimal for the respective players (see von
Neumann and Morgenstern 1964 for explanation of “optimal” and for proof)
and (2) there is an optimal pair of pure strategies among the class of all pairs
of optimal strategies.
Proceeding as in the Model 1 case shows that (a) the expected duration
E(D2 ) of Model 2 is less than or equal to E(DG ); (b) for each sequence of rolls
the Model 2 game ends at the same time or sooner than the corresponding
standard game. Thus if L2 is the least number of turns in any Model 2 game
and LG is the least for G, then L2 ≤ LG . Now count the number of jumps
needed in Model 2 to bear off. Assuming the best case, all rolls (6, 6), the
number of jumps required for a (say) W man at coordinate x is $x/6%, which
242 Edward O. Thorp

gives a total of 5 + 3 · 2 + 5 · 3 + 2 · 4 = 34. Thus LG ≥ L2 = $34/4% = 9 rolls.

It is easy to prescribe sequences of rolls and strategies such that either B or
W can win G in 9 moves. Therefore LG = 9.
We have shown that E(D1 ) ≤ E(D2 (s)) ≤ E(DG (s)). Note that strategies
aren’t relevant for D1 .

Theorem 1. For any strategy pair s, E(D1 ) < E(D2 (s)) < E(DG (s)).

Proof. To show strict inequality, it suffices to give a finite sequence of rolls such
that D1 < D2 (s) for each s and to give a sequence such that D2 (s) < DG (s) for
each s. If (6, 5) is the opening roll and double sixes only are rolled thereafter,
then D1 = 7 and D2 (s) ≥ 9. If (6, 5) is the opening roll and then double fives
only are rolled, D2 (s) = 11 but in G the men on 1, 6, 19, and 24 can never
move so we can make DG (s) ≥ N for arbitrary N . ' (

In both Model 1 and Model 2 the length of the game is bounded above
and the game trees are finite. The proof of the theorem shows this is not the
case for G.

Example 1. A prime (of player X) is 6 or more consecutive cells each “blocked,”

or occupied, by two or more X men. Suppose F and S have all their men in
adjacent primes, each separating the other from his 0 cell. If (6, 6) is rolled
forever by both players, no one moves and the game does not end.

Table 1 gives an alternative “more active” example.

Table 1. Initial configuration: all men borne off except 1 W at 22, 1 B at 3. All
cell numbers are W coordinates. The cycle W 2 · · · B4 is repeated forever.

move no. roll choice

W1 (3, 3) W 22 → 10
B1 (4, 3) B3 → 10, W 10 → 25
W2 (2, 1) W 25 → 22
B2 (2, 3) B10 → 15
W3 (4, 3) W 22 → 15, B15 → 0
B3 (2, 1) B0 → 3
W4 (2, 3) W 15 → 10
B4 (4, 3) B3 → 10, W 10 → 25
W5 (2, 1) W 25 → 22

This leads to the question of whether the expected duration of the game is
finite. Conceivably this could depend on which strategies the players choose.
 Backgammon: The Optimal Strategy for the Pure Running Game 243

Here, but for a gap, is a proof that the expected duration of the game is finite
for arbitrary stationary strategies.2
Assume first that the values of the doubling cube are limited to the finite
set {2k , k = 0, 1, . . . , n}. Then the number of possible states is finite. The
absorbing states are those in which (a) all the men of the player who has just
moved are in his 0 cell and not all the other player’s men are in his zero cell.
The transition probability Pij from state i to state j depends only on the
states i and j. The game is therefore a finite Markov chain with stationary
transition probabilities. We conjecture (this is the gap in the proof) that all
the other states are transient. It follows at once from this that the expected
time to reach an absorbing state, and hence the expected duration of the
game, is finite (e.g., Kemeny and Snell 1960, Theorem 3.1.1, p. 43.).
If instead the possible doubling cube values increase without limit (2k , k =
0, 1, 2, . . .), the same proof works under the additional hypothesis that for
sufficiently large N , the strategy pair is the same for all k ≥ N . Then we
group together all states that have k ≥ N and are otherwise the same. Now
apply the preceding “proof.”

3 The number of backgammon configurations

We shall see that the number of possible configurations is enormous. Thus
any attempt to solve the game by direct calculation would seem to be limited
to parts of the end game. This does not, however, rule out solutions to larger
portions of the game via theorem proving, or by efficient algorithms. Such
solutions might be exact, approximate, or involve inequalities.
Some configurations cannot arise in the course of the game. The rules
specify that no one of cells 1, 2, . . . , 24 can contain both Black and White
men. The rules also specify that the game ends when one player’s men are in
his 0-cell, therefore the configuration where each player has all his men in his
0-cell is impossible.
The impossibility of other configurations may be deduced. For instance,
suppose all men are on the bar. Whoever just moved had to, by the rules,
attempt (successfully in this case) to move a man or men off the bar. This
is a contradiction. As a more complex illustration, suppose both players have
men on the bar and have primes on their own home boards. What happened
on the last move? The player whose move it was had men on the bar and
2 Editor’s note: The gap can be filled. A standard result in Markov chains is that if state
i is recurrent and leads to state j in a finite number of steps with positive probability,
then state j is recurrent and leads back to state i in a finite number of steps with positive
probability. In particular, if every nonabsorbing state leads to an absorbing state in a finite
number of steps with positive probability, then the conjecture that every nonabsorbing
state is transient is correct. So all that needs to be verified is that from any state there
is a finite sequence of dice rolls that leads to one player or the other winning the game,
assuming the specified strategy. This requires a bit of thought but can be verified to be
correct.
244 Edward O. Thorp

attempted to move them. He did not succeed since his opponent had a prime
after the move, hence before. Therefore the configuration did not change. By
induction, the board configuration was always thus, contradicting the initial
configuration specified in the rules. Hence these configurations are impossible.
If we know who has just moved, additional configurations are impossible.
For instance, a player cannot complete his turn with all his men on the bar
when the opponent has no blocked points on his home board.
Some questions about possible configurations: Suppose W has i men on
the bar, B has j men on the bar, and X (= W or B) is to move. Describe
this by (i, j; X). Which (i, j; X) can arise in the course of the game? In par-
ticular what are the maximal such triples (i, j; W ) and (i, j; B)? Presumably
all submaximal triples can also arise.
[Following a suggestion of the referee we find a simple upper bound for the
number of configurations. Using ! the "fact that the number of ways of placing
r balls in n ordered cells is r+n−1 r ! " (Feller 1968), we can place the 15 B
men in the 26 cells {0, . . . , 25} in 4015 ways. The same is true for the 15 W
men. This includes all the possible configurations plus some impossible ones,
such as those with B and W men in the same cell. The resulting number
! "2
M = 40 15 ≈ 1.62 × 1023 is an upper bound for the number of configurations.
Using the rule that opposing men may not occupy the same cell reduces
this somewhat. The bound given in the original paper, via a more complex
combinatorial calculation, was about 1.85 × 1019 .]3
An upper bound for the number of states, assuming a limit of n values for
the doubling cube and initial stakes of one unit, is then 2M n.
We turn now to the important special case of the pure running game.
Definition 5. The pure running game (or end game or no-contact game) is
that portion of the game where no further contact between men is possible, i.e.,
when there exists a real number x such that all W men have coordinates less
than x and all B men have coordinates greater than x. (Here all coordinates
are W coordinates.)
The end game is finite and cannot take more than 63 moves for W , 62 for
B. This case occurs if x = 12.5, all 15 B men are on 13, all 15 W men are
on 12, the only subsequent rolls are (1, 2) and B and W assemble all 15 men
on B1 and W 1 respectively, before bearing off. It follows that the game tree
arising from each initial end-game position is finite.
To get an upper bound K for the number of end-game configurations,
relabel the cells 1, . . . , 26.
3 Editor’s note: A better bound is
15 “ ”“
X 26 15 − 1”“15 + 26 − i − 1”
M = ≈ 1.14 × 1019 .
i=1
i i−1 15

The three factors correspond to the number of ways to choose the i cells occupied by
W , the number of ways to place the 15 W men in these i cells with at least one in each,
and the number of ways to distribute the 15 B men into the remaining 26 − i cells.
 Backgammon: The Optimal Strategy for the Pure Running Game 245

(a) Pick 1 ≤ i < j ≤ 26.

(b) Put 1 W man in cell i, put 14 W men in cells 1, . . . , i.
(c) Put 1 B man in cell j, put 14 B men in cells j, . . . , 26. Then
# # $13 + i%$40 − j %
K= ≈ 1.403 × 1015 .
14 14
1≤i<j≤26

There is one (and possibly other) impossible end-game configuration: all

W men in 0 and all B in 25. As before we believe this upper bound is a good
approximation to the number of possible configurations.
According to the rules, the player can only bear off men (i.e., move them
to his 0-cell) if all men not borne off are in the home! board.
" Thus all 15
W men, e.g., are in cells 0 to 6 which can happen in 21 15 ways. There are
!21"2 !21"2
15 ≈ 2.945 × 10 configurations and 2n 15 ≈ 5.889 × 10 n states, not all
9 9

of which are possible, given n values for the doubling cube.

An interesting special problem is to find the optimal strategy and expec-
tation for those states where each player has exactly one man left to bear off.
[When B and W are out of contact there are 12 · 23 = 276 configurations,
276n positions, and 552n states. In these cases Model 1! and the actual game
coincide and Section 5 below gives the complete solution.]

4 Solution to the Model 1 end game

Label Model 1 states (i, j; W ) with i ≥ 1, j ≥ 0 or (i, j; B) with i ≥ 0, j ≥ 1,
where i is the White goal, j is the Black goal, and W and B designate who is
to move. States (i, 0; W ) and (0, j; B) are terminal states in which B and W ,
respectively, have won. If i, j ≥ 0 then a White move goes from state (i, j; W )
to (max(i−k, 0), j; B) and a Black move goes from state (i, j; B) to (i, max(j −
k, 0); W ), where k is distributed according to the dice probabilities A(k) and
3 ≤ k ≤ 24. For reference 36A(k) is respectively 2, 3, 4, 4, 6, 5, 4, 2, 2, 1 for
k = 3, . . . , 12; 1 for k = 16, 20, 24; and 0 otherwise.
[For the player who is to move next, the expected payoff for the game (as-
suming a unit stake) will be denoted for White by E(i, j; W ) and E(i, j; B).
These will be calculated recursively on a pair of lattices as indicated schemati-
cally in Figures 2 and 3.] Note that the probability P (i, j; X) of White winning
satisfies P (i, j; X) = (E(i, j; X) + 1)/2, where X &= W or B. The two lattices
describe all Model & 1 states. Then E(i, j; W ) = k E(max(i − k, 0), j; B)A(k)
and E(i, j; B) = k E(i, max(j − k, 0); W )A(k) subject to the boundary con-
ditions E(i, 0; W ) = −1 and E(0, j; B) = 1. Equivalently,
i−1
# 24
#
E(i, j; W ) = E(i − k, j; B)A(k) + A(k) (1)
k=3 k=i

and
246 Edward O. Thorp
j−1
# 24
#
E(i, j; B) = E(i, j − k; W )A(k) − A(k), (2)
k=3 k=j

where an empty sum is taken to be zero.

To start the recursion, refer to Figures 2 and 3 and note that for zone
(a1) E(i, j; W ) = 1 for i = 1, 2, 3. Then to find values for zone (b1) we use
(2) to find E(i, j; B) for i = 1, 2, 3. Next to find values for zone (a2) calculate
E(i, j; W ) for i = 4, 5, 6, then for zone (b2) find E(i, j; B) for i = 4, 5, 6, etc.
The matrices of values are readily determined with the aid of a computer.
The matrices will satisfy (error checks)
• |E(i, j; X)| ≤ 1 (the maximum gain or loss is 1)
• E(i, j; W ) = −E(j, i; B) (antisymmetry)
• E(i, j; W ) ≥ E(i, j; B) (it is better for White if White is to move)
• E(i, j; X) ≥ E(i + m, j; X), m ≥ 0 (it is better for White to be closer to
his goal)
• E(i, j; X) ≤ E(i, j +n; X), n ≥ 0 (it is better for White if Black has farther
to go).
The antisymmetry property allows us to replace (1) and (2) by
i−1
# 24
#
E(i, j; W ) = − E(j, i − k; W )A(k) + A(k). (3)
k=3 k=i

Then E(i, j; W ) is calculated recursively by finding (1) the first three rows, (2)
the first three columns, (3) the next three rows, (4) the next three columns,
etc.
We calculated E(i, j; W ) recursively to four places (some roundoff error in
fourth place) by equations (1) and (2) for all i and j such that 1 ≤ i, j ≤ 168.
The calculation is extremely fast. Tables 2 and 3 show results rounded to two
places. Table 4 shows a portion of the full table from which Tables 2 and 3
were extracted. [The rows and columns of Tables 2–4 vary monotonically so
values not given can be obtained by interpolation or extrapolation.]
An alternative to the recursion calculation is to simulate a large number
of Model 1 games on a computer, to determine the matrices of Figures 2 and
3. Equivalently we can determine the matrix (P (i, j; F )) of probabilities that
the first player F will win given that his goal is i and that the goal of the
second player S is j. Such a matrix is computed for selected (i, j) and given
as Table 5. Note that an exact (P (i, j; F )) matrix follows from Table 2 and
the equation P (i, j; F ) = (E(i, j; F ) + 1)/2.
The entries in Table 5 are each based on n = 10,000 games. The stan-
dard
' deviation of the sample mean from the true mean p = P (i, j; F ) is
p(1 − p)/n ≤ 0.005 so the third digit is in doubt. For large (i, j), simulation
is a useful alternative to the recursion method. It provides values for a speci-
fied (i, j) pair without having to calculate values for any, not to say all, (i! , j ! )
pairs with i! ≤ i and j ! ≤ j. Further, P (i! , j; F ) ≥ P (i, j; F ) if i! ≤ i and
 Backgammon: The Optimal Strategy for the Pure Running Game 247

Black coordinates (j)

0 1 2 3 4 5 6 7 8 9

0 ! ! ! ! ! ! ! ! ! !

1 ! ! ! ! ! ! ! ! ! !

! ! ! ! ! ! ! ! ! !
Zone (a1) Value +1
W 2
h
! ! ! ! ! ! ! ! ! !
White wins
i
t 3
e

! ! ! ! ! ! ! ! ! !
c
o 4
o

! ! ! ! ! ! ! ! ! !
r Zone (a2)
d
i 5
n
! ! ! ! ! ! ! ! ! !
a
t 6
e
s
(i) 7 ! ! ! ! ! ! ! ! ! !
!
! ! ! ! ! ! ! ! ! !
Black wins, E(i, 0; W ) = −1
8

9 ! ! ! ! ! ! ! ! ! !

Fig. 2. Lattice of states (i, j; W ), i ≥ 1, j ≥ 0, White to move. Expectation for

White is E(i, j; W ).

P (i, j ! ; F ) ≤ P (i, j; F ) if j ! ≤ j so given P (i! , j ! ; F ) and P (i, j; F ) we have

P (i! , j ! ; F ) ≤ P (r, s; F ) ≤ P (i, j; F ) for i ≤ r ≤ i! and j ! ≤ s ≤ j. Therefore
a table of selected values provides useful information about the remaining
values.
A table equivalent to Tables 2, 3, and 5 appears as Table 1 of Keeler and
Spencer (1975). Their table was computed by simulation and the method was
different than used for our Table 5. However, the agreement with our Table 5
and our exact Tables 2 and 3 is excellent. Their figures are off by at most 1
in the second place.
The first draft of this paper and the paper by Keeler and Spencer were
each written without knowledge of the other. This paper was then revised to
refer to their results.
248 Edward O. Thorp

Black coordinates (j)

0 1 2 3 4 5 6 7 8 9

! ! ! ! ! ! ! ! ! !
White wins, E(0, j; B) = +1
0

1 ! ! ! ! ! ! ! ! ! !

! ! ! ! ! ! ! ! ! !
Zone (b1)
W 2
h
! ! ! ! ! ! ! ! ! !
i
t 3
e

! ! ! ! ! ! ! ! ! !
c
o 4
o

! ! ! ! ! ! ! ! ! !
r Zone (b2)
d
i 5
n
! ! ! ! ! ! ! ! ! !
a
t 6
e
s
(i) 7 ! ! ! ! ! ! ! ! ! !

! ! ! ! ! ! ! ! ! !
Black moves and wins
8

! ! ! ! ! ! ! ! ! !
Value −1
9

Fig. 3. Lattice of states (i, j; B), i ≥ 0, j ≥ 1, Black to move. Expectation for White
is E(i, j; B).

The values in parentheses in Table 5 are those given by Jacoby and Craw-
ford (1970, p. 118). There is a systematic discrepancy between the two sets
of numbers, with the latter set of probabilities noticeably closer to 0.5 than
ours. To “explain” this, note first that Jacoby and Crawford (1970, p. 119) say
that they do not claim their values are “completely accurate” and that they
are based on “experience.” Thus their figures may have substantial errors.
On the other hand, our values are for Model 1 and Model 1 neglects “blots.”
Model 1 also neglects wastage in bearing off. In the real game, with such large
i, j, there will be on average several blots hit on each side. This will have the
effect of increasing the effective i and j, very possibly enough to give values
consistent with those of Jacoby and Crawford.
Table 2. E(i, j; W ) for Model 1 rounded to two decimal places. Decimal points is two places from right and has been omitted.

White Black total

total 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85

5 76 94 98 98 100 100 100 100 100 100 100 100 100 100 100 100 100
10 −36 56 81 90 99 99 100 100 100 100 100 100 100 100 100 100 100
15 −71 −6 40 73 88 94 98 99 100 100 100 100 100 100 100 100 100
20 −85 −59 −11 37 66 82 92 97 99 100 100 100 100 100 100 100 100
25 −97 −79 −51 −10 32 60 78 89 95 98 99 100 100 100 100 100 100
30 −99 −90 −74 −45 −7 29 56 74 86 93 97 99 99 100 100 100 100
35 −100 −97 −87 −68 −41 −7 26 52 71 84 91 96 98 99 100 100 100
40 −100 −99 −94 −83 −64 −37 −6 24 50 68 81 90 94 97 99 99 100
45 −100 −100 −98 −92 −80 −61 −35 −6 23 47 66 79 88 93 96 98 99
50 −100 −100 −99 −97 −90 −77 −58 −33 −5 22 45 63 77 86 92 96 98
55 −100 −100 −100 −98 −95 −87 −74 −55 −31 −5 21 43 61 75 84 91 95
60 −100 −100 −100 −99 −98 −94 −85 −72 −53 −30 −5 20 41 59 73 83 90
65 −100 −100 −100 −100 −99 −97 −92 −83 −69 −51 −28 −4 19 40 57 71 81
70 −100 −100 −100 −100 −100 −99 −96 −90 −81 −67 −49 −27 −4 18 39 56 70
75 −100 −100 −100 −100 −100 −99 −98 −95 −89 −79 −65 −47 −26 −4 18 37 54
80 −100 −100 −100 −100 −100 −100 −99 −97 −94 −87 −77 −63 −46 −25 −4 17 36
85 −100 −100 −100 −100 −100 −100 −100 −99 −97 −93 −86 −76 −62 −44 −25 −4 17
90 −100 −100 −100 −100 −100 −100 −100 −99 −98 −96 −92 −84 −74 −60 −43 −24 −4
95 −100 −100 −100 −100 −100 −100 −100 −100 −99 −98 −95 −90 −83 −72 −59 −42 −23
100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −97 −94 −89 −82 −71 −57 −41
105 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −99 −97 −94 −88 −80 −70 −56
110 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −98 −96 −93 −87 −79 −68
115 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −98 −96 −92 −86 −78
120 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −97 −95 −91 −85
125 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −99 −97 −94 −90
130 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −98 −97 −94
135 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −98 −96
140 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −99 −98
145 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99 −99
150 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −99
155 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100
Backgammon: The Optimal Strategy for the Pure Running Game

160 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100
165 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100 −100
249
 Table 3. Continuation of Table 2.
250

White Black total

total 90 95 100 105 110 115 120 125 130 135 140 145 150 155 160 165

5 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
10 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
15 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
20 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
25 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
30 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
Edward O. Thorp

35 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
40 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
45 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
50 99 99 100 100 100 100 100 100 100 100 100 100 100 100 100 100
55 97 98 99 100 100 100 100 100 100 100 100 100 100 100 100 100
60 94 97 98 99 100 100 100 100 100 100 100 100 100 100 100 100
65 88 93 96 98 99 99 100 100 100 100 100 100 100 100 100 100
70 80 87 92 95 97 98 99 100 100 100 100 100 100 100 100 100
75 68 78 86 91 95 97 98 99 99 100 100 100 100 100 100 100
80 53 67 77 85 90 94 96 98 99 99 100 100 100 100 100 100
85 35 52 65 76 84 89 93 96 98 99 99 100 100 100 100 100
90 16 34 50 64 75 83 88 93 95 97 98 99 99 100 100 100
95 −4 16 34 49 63 73 82 88 92 95 97 98 99 99 100 100
100 −23 −4 15 33 48 62 72 80 87 91 94 96 98 99 99 100
105 −40 −22 −3 15 32 47 60 71 80 86 91 94 96 98 99 99
110 −55 −39 −21 −3 13 31 46 59 70 79 85 90 93 96 97 98
115 −67 −54 −38 −21 −3 14 31 46 58 69 78 84 89 93 95 97
120 −77 −66 −53 −37 −21 −3 14 30 45 57 68 77 84 89 92 95
125 −84 −76 −65 −52 −37 −20 −3 14 30 44 57 67 76 83 88 92
130 −89 −83 −75 −64 −51 −38 −20 −3 13 29 43 56 66 75 82 87
135 −93 −89 −82 −74 −63 −50 −35 −19 −3 13 28 43 55 65 74 81
140 −96 −93 −88 −81 −73 −62 −49 −35 −19 −3 13 28 42 54 65 73
145 −97 −95 −92 −87 −80 −72 −61 −48 −34 −19 −3 13 28 41 53 64
150 −98 −97 −95 −91 −86 −80 −71 −60 −47 −33 −18 −3 12 27 41 53
155 −99 −98 −97 −94 −91 −80 −79 −70 −59 −47 −33 −18 −3 12 27 40
160 −99 −99 −98 −96 −94 −90 −85 −78 −69 −58 −46 −32 −18 −3 12 26
165 −100 −99 −99 −98 −96 −93 −89 −84 −77 −68 −58 −45 −32 −17 −3 12
 Backgammon: The Optimal Strategy for the Pure Running Game 251

Table 4. Part of the four place table of E(i, j; W ) for Model 1. The decimal place
is four places from the right and has been omitted. The complete table (49 pages
like this one) covers all i, j from 1 to 168.

White Black total

total 25 26 27 28 29 30 31 32 33 34 35 36

25 3235 3914 4511 5080 5586 6034 6435 6828 7188 7503 7789 8061
26 2461 3174 3813 4423 4972 5465 5912 6345 6744 7096 7418 7721
27 1676 2414 3086 3732 4323 4860 5353 5828 6267 6657 7015 7350
28 871 1623 2323 3002 3633 4214 4753 5272 5752 6182 6576 6945
29 72 830 1548 2254 2920 3542 4124 4684 5205 5674 6104 6506
30 −706 48 776 1500 2194 2851 3471 4069 4627 5133 5599 6033
31 −1456 −713 15 751 1464 2147 2798 3429 4020 4560 5060 5527
32 −2183 −1456 −733 6 731 1433 2107 2765 3386 3957 4489 4988
33 −2865 −2161 −1450 −718 10 721 1411 2088 2731 3330 3890 4419
34 −3493 −2816 −2126 −1408 −688 23 719 1406 2066 2686 3271 3828
35 −4074 −3427 −2763 −2066 −1363 −661 32 723 1391 2027 2634 3215
36 −4616 −4001 −3365 −2696 −2014 −1329 −644 42 714 1360 1983 2585
37 −5119 −4537 −3933 −3294 −2639 −1974 −1305 −628 41 693 1327 1946
38 −5576 −5029 −4458 −3854 −3229 −2590 −1941 −1279 −617 33 674 1304
39 −6002 −5490 −4953 −4382 −3789 −3179 −2552 −1909 −1260 −615 27 663
40 −6401 −5922 −5418 −4882 −4321 −3739 −3137 −2515 −1882 −1247 −609 27
41 −6775 −6329 −5857 −5352 −4822 −4269 −3693 −3094 −2479 −1857 −1229 −597
42 −7115 −6701 −6260 −5787 −5288 −4764 −4214 −3640 −3047 −2442 −1827 −1205
43 −7429 −7045 −6635 −6192 −5722 −5226 −4704 −4155 −3586 −3001 −2403 −1795
44 −7722 −7367 −6985 −6571 −6129 −5661 −5165 −4642 −4098 −3536 −2957 −2366
45 −7992 −7665 −7310 −6923 −6509 −6067 −5597 −5101 −4581 −4043 −3487 −2915
46 −8233 −7933 −7604 −7245 −6857 −6442 −5999 −5530 −5036 −4523 −3990 −3440
47 −8450 −8176 −7873 −7540 −7179 −6790 −6374 −5931 −5465 −4977 −4469 −3941
48 −8648 −8398 −8121 −7813 −7477 −7115 −6725 −6309 −5869 −5407 −4924 −4419

37 38 39 40 41 42 43 44 45 46 47 48

25 8316 8530 8726 8906 9075 9206 9318 9417 9509 9579 9637 9693
26 8005 8248 8471 8675 8866 9020 9152 9272 9381 9467 9540 9610
27 7663 7935 8184 8414 8628 8805 8960 9102 9230 9336 9426 9511
28 7288 7588 7865 8121 8359 8562 8741 8906 9056 9182 9291 9393
29 6879 7209 7514 7797 8061 8290 8495 8684 8857 9005 9135 9255
30 6437 6798 7132 7442 7733 7989 8221 8436 8633 8805 8957 9096
31 5962 6353 6718 7057 7376 7661 7920 8161 8384 8579 8754 8915
32 5454 5877 6272 6642 6991 7304 7592 7860 8108 8328 8526 8709
33 4917 5372 5799 6199 6577 6920 7236 7531 7805 8050 8273 8478
34 4354 4840 5298 5730 6137 6509 6854 7176 7476 7746 7993 8221
35 3768 4284 4773 5234 5671 6072 6445 6794 7119 7415 7687 7939
36 3163 3706 4223 4714 5179 5608 6009 6385 6736 7058 7354 7630
37 2544 3111 3654 4171 4663 5120 5547 5950 6326 6673 6995 7296
38 1917 2505 3069 3610 4126 4608 5062 5490 5891 6264 6611 6937
39 1286 1889 2472 3033 3570 4075 4553 5006 5432 5830 6202 6553
40 656 1268 1865 2442 2997 3524 4024 4500 4950 5373 5770 6145
41 32 649 1254 1843 2413 2957 3478 3975 4448 4895 5316 5716
42 −581 35 644 1240 1821 2380 2917 3434 3928 4397 4842 5265
43 −1181 −570 38 637 1225 1795 2347 2880 3393 3882 4349 4794
44 −1766 −1164 −561 38 629 1206 1769 2316 2845 3353 3839 4305
45 −2331 −1742 −1148 −554 37 618 1188 1746 2289 2812 3316 3801
46 −2875 −2301 −1719 −1133 −547 34 608 1174 1726 2263 2781 3282
47 −3397 −2841 −2273 −1698 −1119 −541 33 602 1161 1707 2238 2753
48 −3897 −3359 −2808 −2246 −1678 −1106 −535 34 597 1149 1689 2215

The exact calculation for i = j = 167 shows a Model 1 win probability

for F of .5592 (compare Table 5 value .564). This suggests that the value
252 Edward O. Thorp

Table 5. Probabilities of success for first player in Model 1. Each figure based on
10,000 simulated games, assuming player who moves has goal N (F ) and other player
has goal N (S). Standard deviation of error is less than or equal to .005. Parenthetical
numbers are corresponding probabilities given by Jacoby and Crawford (1970).

N (F ) N (S)
50 60 70 80 90 100 110 120 130 140 150 160 167 170 180

50 .610 .817 .930 .977 .995 .999 .999

60 .346 .598 .801 .913 .969 .992 .998
(.545) (.889)
70 .169 .355 .592 .779 .902 .957 .986
(.524) (.857)
80 .068 .190 .365 .582 .764 .883 .953
(.512) (.833)
90 .019 .081 .201 .382 .578 .754 .875
(.508) (.818)
100 .007 .027 .094 .222 .383 .574 .737
(.506) (.800)
110 .001 .008 .037 .104 .225 .399 .568
(.505) (.792)
120 .571 .716 .841 .915 .957 .983 .993
130 .404 .574 .710 .822 .906 .956 .982
140 .259 .412 .564 .713 .822 .901 .955
167 .564

of the actual game for F is positive. Note, however, that in the actual game
E(X) = 7 for the initial roll rather than the 8 1/6 of subsequent rolls, whereas
Tables 2, 3, and 5 assume 8 1/6 for all rolls. (This last is reasonable because
Model 1 is meant to represent the end game.) Examination of the full Tables 2
and 3 suggests that changing the initial roll to conform with the actual game
would reduce the F win probability to about .542, giving a value of .084 to
E.
However, we cannot make useful inferences from this about the value (as-
suming both players choose optimal strategies) of the actual game to F . The
additional features of the actual game make it substantially different from
Model 1. Hitting blots, loss of some or all of a roll when men are sufficiently
blocked, and “wastage” all tend to make N (F ) and N (S) substantially greater
than in Model 1. However, symmetry shows that p = P (i, i; W ) > .50 for
i ≥ 1. Thus we might conjecture that the probability p that F wins satisfies
.55 > p > .50 whence the value V (F ) of the game to F (without doubling or
gammons) would satisfy .10 > V (F ) > 0.
Doubling and gammons both should alter this considerably. If X doubles
and Y folds it is an ordinary win. In scoring the actual game, there are three
cases if no one folds. If X wins and some Y men have reached Y 0 it is an
ordinary win and X receives the value on the cube. If no Y men have reached
Y 0 and all Y men are in cells Y 1 through Y 18 it is a gammon and X gets
twice the value on the cube. If no Y men have reached Y 0 and some Y men
 Backgammon: The Optimal Strategy for the Pure Running Game 253

are in Y 19 through Y 25 it is a backgammon and X receives three times the

value on the cube.
Even without doubling or gammons, V (F ) is not easy to judge. We can’t
even tell if V (F ) > 0. For instance, Cooke and Bradshaw (1974) classify
initial rolls according to how good they judge them to be. There is fairly
general agreement on the approximate classification and as to whether initial
rolls are good or bad. The good rolls for F are 3-1, 4-2, and 6-1 which have
joint probability 6/30, since the initial F roll in the actual game cannot be
doubles. For the second player all first rolls of doubles, except double fives, are
also good to very good, for a probability of 11/36 that S begins with a good
roll. Thus the advantage of the initial roll is offset. But how much? Could the
value of the game even be positive for the second player?

5 Solution to the Model 1! end game

We extend the method to solving Model 1! , which we define as Model 1 ex-
tended to incorporate the doubling cube. The doubling cube is the most in-
teresting, subtle, and least understood aspect of the game. Here for the first
time we will learn things contrary to the beliefs of most players.
Either player may make the initial double. If one player doubles, the other
player may accept or fold. If he accepts, the game continues for doubled stakes
and the player who accepted now possesses the doubling cube, i.e., is the only
player who can make the next (re)double. This must be done just prior to his
roll by the player who possesses the cube. If the doubled or redoubled player
folds, he loses the current stake prior to the double.
The states for Model 1! are described by six lattices, (i, j; X), (i, j; X, D),
and (i, j; X, O), where X = W or B. If the last coordinate is X, no one has
yet doubled and therefore no one has the cube. If the last coordinate is D the
player who is to move has the doubling cube; if it is O the other player has it.
The moves of the game, i.e., the allowable transformations (between lattices),
are:
a. given W
White doubles: (i, j; W ) to (max(i − k, 0), j; B, D)
White does not double: (i, j; W ) to (max(i − k, 0), j; B)
b. given B
Black doubles: (i, j; B) to (i, max(j − k, 0); W, D)
Black does not double: (i, j; B) to (i, max(j − k, 0); W )
c. given W, D
White doubles: (i, j; W, D) to (max(i − k, 0), j; B, D) or
White does not double: (i, j; W, D) to (max(i − k, 0), j; B, O)
d. given W, O
(i, j; W, O) to (max(i − k, 0), j; B, D)
254 Edward O. Thorp

e. given B, D
Black doubles: (i, j; B, D) to (i, max(j − k, 0); W, D) or
Black does not double: (i, j; B, D) to (i, max(j − k, 0); W, O)
f. given B, O
(i, j; B, O) to (i, max(j − k, 0); W, D)
We calculate the expectation for White per unit currently (i.e., at point
i, j) bet. A decision to double is made by comparing the expectation from
doubling, per current unit, with that from not doubling. The expectation, if
positive, is increased by doubling and decreased by the loss of the doubling
cube. Thus whether to double will depend in general on the results of each cal-
culation. In case (a), White’s expectation
& if he doubles is E D ≡ E D (i, j; W ) =
min(1, E ), where E (i, j; W ) = 2 k E(max(i − k, 0), j; B, D)A(k), and his
D D

expectation if he does not double is

#
E 0 (i, j; W ) = E(max(i − k, 0), j; B)A(k).
k

In case (c) White’s expectation if he doubles is

! "
E D ≡ E D (i, j; W, D) = min 1, E D (i, j; W, D) ,

where #
E D (i, j; W, D) = 2 E(max(i − k, 0), j; B, D)A(k),
k

and if he does not double, his expectation is

#
E 0 (i, j; W, D) = E(max(i − k, 0), j; B, O)A(k).
k

We assume each player’s goal is to maximize his own expectation. Then

in cases (a) and (c) White should double if E D > E 0 , in which case we assign
the value E D to E(i, j; W ) or E(i, j; W, D). He is indifferent if E D = E 0 and
he should not double if E D < E 0 . In these cases we set E(i, j; W ) = E 0 . Black
should accept a double if E D = E D < 1, he is indifferent if E D = E D = 1, and
should fold if E D > 1. Let A(i, j) = {m, n} stand for these alternatives, where
m = 0, 1, 2 according to whether White should not double, is indifferent, or
should double, and n = 0, 1, 2 according to whether Black if doubled should
fold, be indifferent, or accept a double. Then for cases (a) and (c) we record
the A(i, j) strategy matrices as the E matrices are calculated.
In case (d) White cannot double so
#
E(i, j; W, O) = E(max(i − k, 0), j; B, D)A(k).
k

The E matrices for Black to move are useful auxiliaries in the recursion but
they can be deduced from those for White to move. This follows, just as in
Model 1, from the analogous antisymmetries for Model 1! :
 Backgammon: The Optimal Strategy for the Pure Running Game 255

E(i, j; W ) = −E(j, i; B), (4)

E(i, j; W, D) = −E(j, i; B, D), (5)
E(i, j; W, O) = −E(j, i; B, O). (6)

For reference, the recursion equations for Black to move are:

Case (b):
E(i, j; B) = min(E 0 , E D ),
where #
E 0 ≡ E 0 (i, j; B) = E(i, max(j − k, 0); W )A(k)
k

and
E D = max(−1, E D )
with #
E D (i, j; B) = 2 E(i, max(j − k, 0); W, D)A(k).
k

Case (e):
E(i, j; B, D) = min(E 0 , E D ),
where #
E 0 ≡ E 0 (i, j; B, D) = E(i, max(j − k, 0); W, O)A(k)
k

and
E D = max(−1, E D )
with #
E D (i, j; B, D) = 2 E(i, max(j − k, 0); W, D)A(k).
k

Case (f):
#
E(i, j; B, O) = 2 E(i, max(j − k, 0); W, D)A(k).
k

It is persuasive that it is at least as good, and generally better, to possess

the doubling cube than not to. This ought to be true in the standard game as
well as in the various simpler models we discuss. This is generally taken for
granted by players but a formal proof remains to be given.
We give a proof for Model 1! , in which case it is equivalent to the simul-
taneous inequalities

E(i, j; W, D) ≥ E(i, j; W ) ≥ E(i, j; W, O) (7)

E(i, j; B, D) ≥ E(i, j; B) ≥ E(i, j; B, O) (8)

Theorem 2. Inequalities (7) and (8) hold for Model 1! , i.e., it is always at
least as good to have the doubling cube as not to.
256 Edward O. Thorp

Proof. For i = 1, 2, 3 all parts of (7) are 1. This corresponds to zone (a1) of Fig-
ure 2, in each of the lattices {(i, j; W, D)}, {(i, j; W )}, and {(i, j; W, O)}. Now
let i = 1, 2, 3 and suppose Black follows the (not necessarily optimal) strat-
egy of not doubling when& in states (i, j; B, D) or (i, j; B). Then& his expecta-
tion is, respectively, − k E(i, max(j −k, 0); W, O)A(k) or − k E(i, max(j −
0); W )A(k), whereas the expectation of a Black player in state (i, j; B, O) is
k, &
− k E(i, max(j − k, 0); W, D)A(k). Comparing term by term and using what
was proved about “zone (a1)” establishes (8) for i = 1, 2, 3, i.e., for the regions
of (i, j; B, y), y = D or blank or O, corresponding to zone (b1) of Figure 3.
We continue with zone (a2), zone (b2), etc. just as in the original recursion.
This concludes the proof. ( '
The player who is doubled should accept the double if as a result his
expectation E > −1, he is indifferent if E = −1, and he should reject the
double if E < −1. In this last instance the game ends and the other player
receives +1 so this value, not −E, is assigned to the appropriate state for
use in subsequent recursions. Thus for each state in each of the matrices
{(i, j; W, D)} and {(i, j; B, D)} we compute E D and E 0 . Then we will list
whether the player doubles (−1 = NO, 0 = INDIFF, +1 = YES), whether
the doubled player accepts (−1 = NO, 0 = INDIFF, +1 = YES), and the
expectation E.
Tables 6 and 7 are selected portions of four-place tables for E(i, j; W, D)
and E(i, j; W, O), respectively. [A comparison of corresponding entries in the
two tables shows the enormous value of the doubling cube.]
As a check, the various Model 1! tables were shown to satisfy (4), (5), (6),
(7), (8) as well as monotonicity:
E(i, j; X, Y ) ≥ E(i + k, j; X, Y ) (9)
E(i, j; X, Y ) ≤ E(i, j + k; X, Y ). (10)
and that a player with three or less steps wins on the next move:
E(i, j; W, Y ) = 1 if i = 1, 2, 3 (11)
E(i, j; B, Y ) = −1 if j = 1, 2, 3. (12)
As a further test we have:
Lemma 1. If 0 < E(i, j; W, O) < 1/2, then E(i, j; W, D) ≥ 2E(i, j; W, O)
and W should double.
Proof. If W cannot double, the next state will be of the form (m, n; B, D). If
W can and does double, the next state will also be of this form with the same
probabilities of m, n. Thus W gets 0 < E(i, j; W, O) < 2E(i, j; W, O) < 1 by
doubling.
However (W, D) can choose not to double in which case the next state
is (m, n; B, O). It can and does happen that this sometimes gains more (see
Tables 6 and 7 for many (i, j) examples) hence E(i, j; W, D) > 2E(i, j; W, O)
may occur. ' (
 Backgammon: The Optimal Strategy for the Pure Running Game 257

The optimal strategies were computed directly as previously described.

However an optimal strategy can be derived from the E tables. To illustrate,
we do this using Tables 6 and 7 for the case when White has the doubling cube
and is to move. The procedure is to analyze cases (let E(i, j; W, D) = ED ,
E(i, j; W, O) = EO ), as in Table 8.
Applying these cases to Table 6 produces the shaded area. It also is another
check on the tables.
Keeler and Spencer (1975) approach the Model 1! problem by replacing
the actual dice with a random variable having values 5, 10, and 20 with
probabilities .55, .36, and .09. (The text suggests that their Model 1 simulation
also used this simplification. However the difference in probabilities between
Table 6 and 7 suggests otherwise.) These probabilities are “. . . set to make
the mean = 8 1/6 and the variance = 19, which are the real backgammon
values.”
& Actually the correct variance is 18.472. This comes from the formula
( i x2i − nx2 )/n.
& The incorrect value 19 arises for the actual dice if one uses
the formula ( i x2i − nx2 )/(n − 1), which is the appropriate formula for the
different situation of independent sampling from a population whose mean is
unknown.
Replacing the actual dice by the “5, 10, 20” dice greatly simplifies the
problem: only (i, j) multiples of 5 need to be considered. Thus 170 × 170
matrices become 34 × 34 matrices. The matrix storage sizes are reduced by a
factor of 25. Also the recursion to calculate each entry involves sums of only
three terms, rather than 14 so we have another reduction by a factor of almost
5. Thus the total calculation is reduced by a factor of more than 100. Yet the
Keeler–Spencer results are an excellent approximation.
For instance the probabilities for equal counts in their Table 2 are off by
at most one in the second place. When the opponent has 5 more points they
are within 2 in the second place (all the errors here are overestimates). When
the opponent has 10 more points their figures overestimate by 1 to 4 in the
second place.
The correct Model 1! strategy for doubling, redoubling, and folding, has a
somewhat irregular pattern. Gillogly and Keeler (1975) give a simplified rule
based, presumably, on the Keeler–Spencer calculations:
Approximate rule (Gillogly–Keeler): With pip count C let D = C if C ≤ 25
and D = 1.1C if C > 25. Then make the first double if the opponent’s count
j ≥ D − 2. Redouble if j ≥ D − 1. The opponent should fold if (doubled and)
j ≥ D + 2.
Prior to this paper, Keeler and Spencer (1975), and Gillogly and Keeler
(1975), the rule recommended by experts is the one given by Jacoby and
Crawford (1970):
Approximate rule (Jacoby–Crawford): With a pip count of C double if
j ≥ 1.075C; redouble if j ≥ 1.1C; the opponent should fold if j ≥ 1.15C.
Obolensky and James (1969) gives no rule; Lawrence (1973) and Genud
(1975) repeat the rule without acknowledgment (a common practice in gaming
circles). Stern (1973) gives a complex and unclear set of maxims.
Table 6. Model 1! selected portion of table of E(i, j; W, D) values (White expectation, White to move, White has doubling cube).
258

White should double if and only if (i, j) is within or above the shaded area. Black should accept unless (i, j) is above the shaded area,
in which case he should fold.

White Black total

total 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45

30 5068 5897 7001 8234 9354 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000
Edward O. Thorp

31 4130 5018 5860 6850 8060 9168 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000
32 3180 4096 4991 5819 6680 7872 8971 9985 10000 10000 10000 10000 10000 10000 10000 10000
33 2253 3170 4083 4955 5761 6499 7679 8774 9778 10000 10000 10000 10000 10000 10000 10000
34 1367 2266 3171 4057 4901 5693 6423 7494 8576 9594 10000 10000 10000 10000 10000 10000
35 526 1396 2279 3156 4011 4837 5631 6352 7309 8401 9427 10000 10000 10000 10000 10000
36 −273 563 1418 2274 3123 3959 4775 5556 6282 7147 8245 9277 10000 10000 10000 10000
37 −1028 −227 595 1427 2257 3088 3913 4723 5495 6223 7004 8102 9121 10000 10000 10000
38 −1736 −972 −183 618 1427 2243 3063 3879 4675 5445 6171 6871 7953 8969 9927 10000
39 −2403 −1676 −920 −148 635 1431 2237 3046 3846 4635 5399 6120 6780 7805 8820 9777
40 −3030 −2339 −1617 −873 −115 657 1444 2239 3031 3821 4599 5352 6062 6720 7660 8673
41 −3620 −2961 −2272 −1558 −826 −76 688 1464 2242 3021 3797 4561 5300 6003 6660 7518
42 −4168 −3543 −2885 −2202 −1498 −774 −32 722 1481 2245 3011 3772 4519 5248 5946 6601
43 −4680 −4087 −3462 −2808 −2134 −1437 −721 10 749 1495 2246 2997 3742 4477 5198 5892
44 −5164 −4600 −4004 −3382 −2736 −2068 −1378 −672 45 772 1506 2243 2978 3711 4439 5151
45 −5621 −5083 −4516 −3923 −3306 −2665 −2003 −1323 −629 78 794 1515 2238 2962 3686 4403
46 −6047 −5539 −4999 −4433 −3845 −3233 −2598 −1943 −1274 −588 109 814 1523 2236 2951 3664
47 −6443 −5965 −5455 −4918 −4357 −3773 −3165 −2537 −1890 −1227 −550 138 833 1534 2237 2942
48 −6812 −6365 −5886 −5379 −4845 −4287 −3706 −3104 −2482 −1840 −1183 −512 167 854 1546 2239
Table 7. Model 1! selected portions of four place table of E(i, j; W, O) values (White expectation when White is to move and Black
has the doubling cube). The last digit in this table and in Table 6 may be off by one because of machine round-off errors.

White Black total

total 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45

30 2088 2819 3500 4117 4677 5190 5659 6091 6486 6858 7200 7519 7802 8063 8303 8520
31 1246 2028 2760 3425 4030 4584 5093 5561 5990 6393 6768 7118 7433 7725 7996 8242
32 374 1201 1980 2691 3340 3936 4485 4993 5457 5894 6302 6685 7032 7355 7656 7933
33 −511 351 1169 1922 2612 3249 3840 4387 4889 5361 5804 6219 6599 6953 7285 7592
34 −1396 −513 336 1124 1853 2529 3160 3747 4288 4797 5274 5723 6136 6521 6883 7220
35 −2292 −1385 −518 299 1063 1777 2447 3074 3654 4201 4714 5196 5641 6057 6449 6815
36 −3212 −2268 −1383 −546 248 997 1705 2370 2990 3573 4122 4638 5116 5563 5984 6378
37 −4157 −3174 −2253 −1400 −584 196 938 1641 2298 2918 3502 4051 4561 5038 5488 5911
38 −5094 −4097 −3138 −2249 −1420 −618 155 891 1583 2238 2855 3436 3977 4484 4963 5414
39 −5994 −5029 −4043 −3112 −2254 −1437 −641 123 846 1533 2182 2793 3365 3902 4410 4889
40 −6830 −5923 −4965 −3993 −3096 −2253 −1443 −658 92 807 1485 2125 2727 3294 3830 4336
41 −7570 −6760 −5848 −4889 −3951 −3072 −2243 −1445 −674 66 769 1436 2065 2661 3225 3759
42 −8169 −7499 −6679 −5751 −4822 −3903 −3042 −2230 −1447 −688 38 729 1384 2006 2598 3159
43 −8605 −8117 −7420 −6574 −5672 −4751 −3854 −3013 −2221 −1451 −706 6 685 1333 1951 2537
44 −8911 −8571 −8060 −7318 −6486 −5594 −4686 −3811 −2991 −2215 −1459 −728 −28 643 1285 1897
45 −9119 −8892 −8531 −7968 −7231 −6400 −5521 −4626 −3773 −2974 −2212 −1469 −751 −59 605 1240
46 −9261 −9106 −8866 −8462 −7886 −7147 −6322 −5455 −4571 −3741 −2960 −2211 −1480 −771 −86 570
47 −9352 −9255 −9089 −8819 −8403 −7809 −7072 −6250 −5390 −4524 −3714 −2948 −2210 −1487 −786 −111
48 −9435 −9352 −9250 −9065 −8780 −8349 −7742 −7004 −6178 −5328 −4482 −3687 −2935 −2206 −1493 −802
Backgammon: The Optimal Strategy for the Pure Running Game
259
260 Edward O. Thorp

Table 8. Analysis of cases.

case optimal strategy value of ED

EO < 0 do not double ≥ EO

EO < 0, ED $= 0 do not double >0
EO = 0, ED = 0 any 0
1
2
> EO > 0, ED = 2EO double; B should accept 1 > ED = 2EO > 0
1
2
> EO > 0, ED > 2EO do not double 1 ≥ ED > 2EO > 0
EO = 12 double; not doubling∗ may 1
or may not also be optimal;
folding∗ is optimal for
Black; accepting may or
may not also be.
EO > 1
2
double; not doubling∗ may 1
or may not also be optimal;
B should fold if doubled.
∗
Doubling gives the opponent the chance to err; not doubling does not, so if
both are optimal doubling is recommended.

Both the J–C and G–K strategies are quite close to the exact results for
Model 1! . Note that in the actual game the effective i and j will tend to
be greater than the pip counts, due to “wastage” in bearing off. This tends
to slightly increase the double, redouble and fold numbers in Table 5.3. We
expect subsequent more realistic calculations to give an increase of 1 just
before some of the change-over points. For instance, when W has 43 White
might redouble at 5 instead of 4 and Black might wait until 8 to fold instead
of 7. Of course the use of the corrected count (Section 6, below) is a more
precise way to incorporate this wastage effect.
Table 9 compares the approximate rules with the exact results if 1 ≤
C ≤ 65, for the case when White redoubles and Black considers accepting or
folding. This region of small C is the most irregular, and even here the G–K
approximate rule is generally good.
For fixed White count i, let Di be the least Black count j for which White
should double. Let Ri be the least Black count for which White should re-
double, and let Fi be the least Black count for which White should fold. It is
plausible to expect the Di ≤ Ri ≤ Fi for all i. The last inequality corresponds
to the obvious fact that there is nothing (i.e., no expectation) to be gained by
folding without being redoubled. Similarly we prove Di ≤ Fi . Note however
that if a player’s expectation is −1, whence he has no probability of winning
he may as well fold voluntarily to save time unless his opponent is a far from
optimal player.
This establishes the first part of the following theorem.
 Backgammon: The Optimal Strategy for the Pure Running Game 261

Table 9. Comparison of Gillogly–Keeler approximate rule and exact rule, for re-
doubling (R) and folding (F ), if 1 ≤ C ≤ 65.

W has W : R if B has B: F if B has

Exact G–K J–C Exact G–K J–C
1–6 Any −1 1 Any 2 1
7 −6 −1 1 1 2 2
8 −1 −1 1 1 2 2
9 −1 −1 1 0 2 2
10 −2 −1 1 0 2 2
11–12 −2 −1 2 0 2 2
13 −2 −1 2 1 2 2
14–18 −1 −1 2 2 2 3
19 −1 −1 2 3 2 3
20 0 −1 2 3 2 3
21–22 0 −1 3 3 2 4
23–25 1 −1 3 4 2 4
26 1 2 3 4 5 4
27 1 2 3 5 5 5
28–30 2 2 3 5 5 5
31 2 3 4 5 6 5
32–33 2 3 4 6 6 5
34–37 3 3 4 6 6 6
38 3 3 4 7 6 6
39–40 4 3 4 7 6 6
41–43 4 4 5 7 7 7
44–46 5 4 5 8 7 7
47–50 5 4 5 8 7 8
51 5 5 6 8 8 8
52–53 6 5 6 9 8 8
54–59 6 5 6 9 8 9
60 6 5 6 10 8 9
61–65 7 6 7 10 9 10

Theorem 3. It is an optimal strategy to fold only when the expectation would

be less than −1 by not folding. With this folding strategy and any optimal
redoubling strategy,
Ri ≤ Fi for all i.
The inequality Di ≤ Ri holds for all i for either of the following optimal
strategy cases:
(a) White redoubles only when he gains by doing so (i.e., he does not
redouble if he is indifferent). He doubles whenever he gains and he doubles or
not, as he pleases, when he is indifferent.
262 Edward O. Thorp

(b) White redoubles whenever he gains and does as he pleases when he is

indifferent. He doubles whenever he gains or is indifferent.
Proof. Note first that the qualifications (a) or (b) are needed because in gen-
eral the optimal strategies are not unique. For doubling, redoubling, and fold-
ing there are lower and upper numbers, Di ≤ Di , Ri ≤ Ri , and F i ≤ F i ,
such that the player is indifferent if and only if Di ≤ j < Di , etc. If these
indifference zones overlap, there exist optimal strategies which violate one or
both of the inequalities Di ≤ Ri ≤ Fi . The proof that Di ≤ Ri uses the
conditions of cases (a) and (b) to rule out these overlaps.
To establish that Di ≤ Ri under conditions (a) or (b) it suffices to prove
Di ≤ Ri and Di ≤ Ri . To prove Di ≤ Ri , note that it is equivalent to showing
that ! "
min 1, E D (i, j; W, D) > E 0 (i, j; W, D)
implies ! "
min 1, E D (i, j; W ) > E 0 (i, j; W ).
This is equivalent to asserting that
( )
#
min 1, 2 E(max(i − k, 0), j; B, D)A(k)
k
#
> E(max(i − k, 0), j; B, O)A(k)
k

implies
( )
#
min 1, 2 E(max(i − k, 0), j; B, D)A(k)
k
#
> E(max(i − k, 0), j; B)A(k)
k

or equivalently that
# #
2 E(max(i − k, 0), j; B, D)A(k) > E(max(i − k, 0), j; B, O)A(k)
k k

implies
# #
2 E(max(i − k, 0), j; B, D)A(k) > E(max(i − k, 0), j; B)A(k).
k k

But this is equivalent to

# #
E(max(i − k, 0), j; B, D)A(k) ≥ E(max(i − k, 0), j; B)A(k).
k k

This follows from E(i, j; B, O) ≥ E(i, j; B) for all i, j, which in turn follows
from (8). Thus Di ≤ Ri .
To show Di ≤ Ri we repeat the proof using ≥ throughout in place of >.
This concludes the proof of the theorem. ' (
 Backgammon: The Optimal Strategy for the Pure Running Game 263

An ambiguity that appears frequently in the previous work (e.g., in Ja-

coby and Crawford 1970, Gillogly and Keeler 1975) is to give the doubling,
redoubling, and folding thresholds in terms of the probability of winning by
the player who is to move. For example Gillogly and Keeler say that the prob-
ability of winning with i = j = 30 is .65, and that the cutoffs for D, R, F are
respectively probabilities of .65, .69, 75. It seems clear from their text and
from comparison with our results that they mean Model 1 probabilities. Note
that the Model 1! probabilities of winning with optimal strategies will in gen-
eral be different, and may even be multi-valued because they depend on the
player’s choice of strategy, and even the optimal strategies may be nonunique.
In general, the probability of winning may depend strongly on strategy. As an
extreme example to make the point, suppose in an actual game both Black
and White always double. White always accepts if doubled, but Black always
folds. Then the probability is one that White wins.
As an instance from the actual game, suppose White has only one man
left, on W 6 and Black has only one man left, on B3. White is to move and
has the cube. Then E(6, 3; W ) = 1/2 (Model 1) because the probability W
wins in the next roll is 3/4. Otherwise Black wins.
White should double, giving E D = 1. Then if Black plays optimally, he
is indifferent as to whether he accepts or folds because in either case his ex-
pectation is −1. Suppose Black (by means of a chance device) accepts with
probability p and folds with probability 1 − p. Then he wins the game with
probability p/4, and White wins with probability 1 − p/4, 0 ≤ p ≤ 1. Hence,
even though both players choose an optimal strategy, the White win proba-
bility may assume any value from .75 to 1 depending on the Black choice of
optimal strategy. It follows that (a) optimal strategies are nonunique in all
the versions of the game which we consider, (b) for any position which can
lead to the above position, this is also true, and (c) in all versions of the game
that we consider, for any position that can lead to the above position (i.e., in
“most” positions), the win probability is not uniquely determined even if the
players limit themselves to optimal strategies.

6 Pending research

In a future paper we plan to present exact strategies for bearing off. With these
we shall extend the results to Model 2, Model 2! (= Model 2 plus doubling
cube), Model 3 (Model 2 with bearing off allowed only when all men have
reached the inner board), and Model 3! (Model 3 with the doubling cube).
Model 3! is the pure running game with the actual rules. The solutions involve
relatively minor corrections or “perturbations” of the Model 1 and Model 1!
solutions. For instance, Gillogly and Keeler (1975) give approximate rules for
estimating the “wastage” of an inner board position. This is the amount that
should be added to the (Model 1, 1! ) pip count i due to the fact that in Models
3 and 3! some parts of the dice counts may be wasted in bearing off.
264 Edward O. Thorp

Approximate rule for wastage (Gillogly–Keeler): Suppose all n men are on

the inner board and cover m points. Let k men be on the one point. Then
add w to i where w = k + (n − 1) + (n − m).
Thus there are penalties for men on the one point, for extra men, and for
extra men clustered on a point.
The rule is not used consistently by its authors. In one example they
compare 5 men on the 6 point with 10 men on the one point. The wastage is
8 for the 5 men on the 6 point and 28 for the 10 men on the one point. The
adjusted count should be i = j = 38. But in the text only the difference of 20
is added to j, giving i = j = 30. In our opinion, their “netting out” method
is not a correct application of their rule.
The methods of this paper apply mutatis mutandis to numerous variants
of the game (Bell 1969). Some earlier forms of the game coincide with Models
1 or 1! and are thus completely solved.
I wish to thank Steven Mizusawa for his programming of the computations
in this paper.
[Author’s afterword: This article was reported on in Thompson (1975)
and Thorp (1975). The exact solution to a special case of Model 3! , the
pure running game with the doubling cube, where each player has borne
off all but just one or two men in their home boards, appeared in Thorp
(1978a,b,c), and was reprinted in Thorp (1984). The exact Model 3! solu-
tion to pure races when each player is bearing off from his home board has
since been computed for hundreds of millions of cases, including up to 9
men for each player, by Hugh Sconyers, a two-time world team backgammon
player. See http://www.back-gammon.info/sconyers/hypergammon.html.
Zadeh and Kobliska (1977) published an important article on optimal dou-
bling in backgammon. Thorp (1977) covers the relation between their paper
and this article, as well as unpublished work by Gillogly, Kahn and Smolen.
The original version of this article has been modified slightly for publi-
cation here. I have incorporated suggestions from the referee and the editor,
as well as some changes of my own. These preserve the historical sense and
smooth the reader’s path. The longer revisions are set off in brackets.]

References
1. Bell, R. C. (1960) Board and Table Games from Many Civilizations, Vol. 1.
Oxford University Press, London.
2. Bell, R. C. (1969) Board and Table Games from Many Civilizations, Vol. 2.
Oxford University Press, London.
3. Cooke, Barclay and Bradshaw, Jon (1974) Backgammon: The Cruelest Game.
Random House.
4. Feller, William (1968) An Introduction to Probability and Its Applications, Vol.
1, third edition. Wiley, New York.
 Backgammon: The Optimal Strategy for the Pure Running Game 265

5. Genud, Lee (1974) Lee Genud’s Backgammon Book. Cliff House Books, Los
Angeles.
6. Gillogly, J. J. and Keeler, E. B. (1975) Playing the running game in backgam-
mon. Popular Bridge 9 (3) 34–38.
7. Holland, Tim (1973) Beginning Backgammon. McKay, New York. Tartan pa-
perback, 1974.
8. Hopper, Millard (1972) Win at Backgammon. Dover. Unabridged reproduction
of the 1941 original.
9. Jacoby, Oswald and Crawford, John R. (1970) The Backgammon Book. The
Viking Press, New York.
10. Keeler, Emmett B. and Spencer, Joel (1969) Proper raising points in a general-
ization of backgammon. The Rand Corporation, 4078 (May).
11. Keeler, Emmett B. and Spencer, Joel (1975) Optimal doubling in backgammon.
Operations Research 23 (6) (Nov.–Dec.) 1063–1071. [Reprinted in Levy (1988).]
12. Kemeny, John G. and Snell, J. Laurie (1960) Finite Markov Chains. Van Nos-
trand, Princeton, NJ.
13. Lawrence, Michael S. (1973) Winning Backgammon. Pinnacle Books, New York.
14. [Levy, David N. L., editor (1988) Computer Games I. Springer-Verlag, New
York.]
15. Neumann, J. von and Morgenstern, Oskar (1964) Theory of Games and Eco-
nomic Behavior. Wiley (Science Editions).
16. Obolensky, Prince Alexis and James, Ted (1969) Backgammon, the Action
Game. Macmillan, Collier Books.
17. Stern, Don (1973) Backgammon, the Quick Course to Winning Play. Corner-
stone Library, Simon and Schuster distributors, Drake Publishers, New York.
18. Thomsen, Dietrick (1975) Beating the game. Science News 107 (12) (Mar. 22)
198–199.
19. Thorp, Edward O. (1975) Recent results and open questions for some particular
games. Preliminary Report. Paper delivered at American Math. Soc. Annual
Meeting, Washington, D.C., Jan. 1975. See Notices, A.M.S., Jan. 1975, Abstract
No. 720-90-7, page A-254.
20. [Thorp, Edward O. (1977) Review of Zadeh and Kobliska. Mathematical Reviews
H449 362.]
21. [Thorp, Edward O. (1978a) End positions in backgammon. Gambling Times
(Sept.). Reprinted in Gammon Magazine preview issue. Reprinted in Levy
(1988).]
22. [Thorp, Edward O. (1978b) End positions in backgammon II. Gambling Times
(Oct.). Reprinted in Levy (1988).]
23. [Thorp, Edward O. (1978c) End positions in backgammon III. Gambling Times
(Nov.). Reprinted in Levy (1988).]
24. [Thorp, Edward O. (1984) The Mathematics of Gambling. Gambling Times,
Incorporated.]
25. [Zadeh, Norman and Kobliska, Gary (1977) On optimal doubling in backgam-
mon. Management Science 23 (8) (April) 853–858. Reprinted in Levy (1988).]