NON-COOPERATIVE GAMES

A DISSERTATION

Presented to the Faculty of Princeton University in Candidacy for

the Degree of Doctor of Philosophy

Recommended for Acceptance by the Department of Mathematics

May, 1950
 Abstract

This paper introduces the concept of a non-cooperative game and develops

methods for the mathematical analysis of such games. The games considered
are n-person games represented by means of pure strategies and pay-off func-
tions defined for the combinations of pure strategies.

The distinction between cooperative and non-cooperative games is unre-

lated to the mathematical description by means of pure strategies and pay-off
functions of a game. Rather, it depends on the possibility or impossibility of
coalitions, communication, and side-payments.

The concepts of an equilibrium point, a solution, a strong solution, a sub-

solution, and values are introduced by mathematical definitions. And in later
sections the interpretation of these concepts in non-cooperative games is dis-
cussed.

The main mathematical result is the proof of the existence in any game of
at least one equilibrium point. Other results concern the geometrical struc-
ture of the set of equilibrium points of a game with a solution, the geometry
of sub-solutions, and the existence of a symmetrical equilibrium point in a
symmetrical game.

As an illustration of the possibilities for application a treatment of a simple

three-man poker model is included.
Contents

1 Introduction 1

2 Formal Definitions and Terminology 2

3 Existence of Equilibrium Points 5

4 Symmetries of Games 7

5 Solutions 9

6 Geometrical Form of Solutions 13

7 Dominance and Contradiction Methods 15

8 A Three-Man Poker Game 17

9 Motivation and Interpretation 21

10 Applications 25

Bibliography 27

11 Acknowledgements 27
1 Introduction

Von Neumann and Morgenstern have developed a very fruitful theory

of two-person zero-sum games in their book:
Theory of Games and Economic Behavior.
This book also contains a theory of n-person games of a type which
we would call cooperative. This theory is based on an analysis of the
interrelationships of the various coalitions which can be formed by the
players of the game.
Our theory, in contradistinction, is based on the absence of coali-
tions in that it is assumed that each participant sets independently,
without collaboration or communication with any of the others.
The notion of an equilibrium point is the basic ingredient in our
theory. This notion yields a generalization of the concept of the so-
lution of a two-person zero-sum game. It turns out that the set of
equilibrium points of a two-person zero-sum game is simply the set of
all pairs of opposing ''good strategies''.
In the immediately following sections we shall define equilibrium
points and prove that a finite non-cooperative game always has at
least one equilibrium point. We shall also introduce the notions of
solvability and strong solvability of a non-cooperative game and prove
a theorem on the geometrical structure of the set of equilibrium points
of a solvable game.
As an example of the application of our theory we include a solution
of a simplified three person poker game.
The motivation and interpretation of the mathematical concepts
employed in the theory are reserved for discussion in a special section
of this paper.

1
2 Formal Definitions and Terminology

In this section we define the basic concepts of this paper and set
up standard terminology and notation. Important definitions will
be preceded by a sub-title indicating the concept defined. The non-
cooperative idea will be implicit, rather than explicit, below.

Finite Game:

For us an n-person game will be a set of n players, or positions,

each with an associated finite set of pure strategies; and correspond-
ing to each player, i, a pay-off function, pi , which maps the set of all
n-tuples of pure strategies into the real numbers. When we use the
term n-tuple we shall always mean a set of n items, with each item
associated with a different player.

Mixed Strategy, si :

A mixed strategy of player i will be a collection of non-negative

numbers which have unit sum and are in one to one correspondence
with his pure strategies.
We write si = ciα πiα with ciα = 1 and ciα ≥ 0 to represent
P P
α α
such a mixed strategy, where the πiα ’s are the pure strategies of player
i. We regard the si ’s as points in a simplex whose vertices are the
πiα ’s. This simplex may be regarded as a convex subset of a real
vector space, giving us a natural process of linear combination for the
mixed strategies.
We shall use the suffixes i, j, k for players and α, β, γ to indicate
various pure strategies of a player. The symbols si , ti and γi , etc. will

2
indicate mixed strategies; πiα will indicate the ith player’s αth pure
strategy, etc.

Pay-off function, pi :

The pay-off function, pi , used in the definition of a finite game

above, has a unique extension to the n-tuples of mixed strategies which
is linear in the mixed strategy of each player [n-linear]. This extension
we shall also denote by pi , writing pi (s1 , s2 , ..., sn ).
We shall write s or t to denote an n-tuple of mixed strategies and
if s = (s1 , ..., sn ) then pi ( s ) shall mean pi (s1 , ..., sn ). Such an n-tuple,
s , will also be regarded as a point in a vector space, which space could
be obtained by multiplying together the vector spaces containing the
mixed strategies. And the set of all such n-tuples forms, of course, a
convex polytope, the product of the simplices representing the mixed
strategies.
For convenience we introduce the substitution notation ( s ; ti ) to
stand for (s1 , s2 , ..., si−1 , ti , si+1 , ..., sn ) where s = (s1 , s2 , ..., sn ). The
effect of successive substitutions (( s ; ti ); γj ) we indicate by ( s ; ti ; γj ),
etc.

Equilibrium Point:

An n-tuple s is an equilibrium point if and only if for every i

pi ( s ) = max[pi ( s ; ti )] (1)
∀γj

Thus an equilibrium point is an n-tuple s such that each player’s

mixed strategy maximizes his pay-off if the strategies of the others are
held fixed. Thus each player’s strategy is optimal against those of the

3
others. We shall occasionally abbreviate equilibrium point by eq. pt.

We say that a mixed strategy si uses a pure strategy πiβ if si =

ciα πiα and ciβ > 0. If s = (s1 , s2 , ..., sn ) and si uses πiα we also say
P
α
that s uses πiα .
From the linearity of pi (s1 , ..., sn ) in si :

max[pi ( s ; γi )] = max[pi ( s ; πiα )] (2)

∀γi α

We define piα ( s ) = pi ( s ; πiα ).

Then we obtain the following trivial necessary and sufficient condition
for s to be an equilibrium point:

pi ( s ) = max piα ( s ) (3)

α

If s = (s1 , s2 , ..., sn ) and si := ciα πiα then

P
α
ciα piα ( s ), consequently for (3) to hold we must have ciα = 0
P
pi (s) =
α
whenever piα ( s ) < max piβ ( s ), which is to say that s does not use
β
πiα unless it is an optimal pure strategy for player i. So we write

if πiα is used in s then piα ( s ) = max piβ ( s ) (4)

β

as another necessary and sufficient condition for an equilibrium

point.
Since a criterion (3) for an eq. pt. can be expressed as the equating
of two continuous functions on the space of n-tuples s the eq. pts.
obviously form a closed subset of this space. Actually, this subset is
formed from a number of pieces of algebraic varieties, out by other
algebraic varieties .

4
3 Existence of Equilibrium Points

I have previously published [Proc. N. A. S. 36 (1950) 48-49] a proof of

the result below based on Kakutani’s generalized fixed point theorem.
The proof given here uses the Brouwer theorem.

The method is to set up a sequence of continuous mappings:

s → s ( s , 1); s → s ( s , 2);...
whose fixed points have an equilibrium point as limit point. A limit
mapping exists , but is discontinuous, and need not have any fixed
points.

Theorem 1. Every finite game has an equilibrium point.

Proof. Using our standard notation, let s be an n-tuple of mixed
strategies, and piα ( s ) the pay-off to player i if he uses his pure strat-
egy πiα and the others use their respective mixed strategies in s . For
each integer λ we define the following continuous functions of s :

qi ( s ) = max piα ( s ),
α
1
φiα ( s , λ) = piα − qi ( s ) + , and
λ
φ+
iα ( s , λ) = max[0, φiα ( s , λ)].

Now φ+ + 1
> 0 so that:
P
iα ( s , λ) ≥ max φiα ( s , λ) =
α λ
α
φ+ ( s , λ)
c0iα ( s , λ) = P iα+ is continuous.
φiβ ( s , λ)
β

Define s0i (s, λ) = πiα c0iα (s, λ) and s(s, λ) = (s01 , s02 , ..., s0n ). Since
P
α
all the operations have preserved continuity, the mapping s → s (s, λ)

5
is continuous; and since the space of n-tuples,
s , is a cell, there must be a fixed point for each λ. Hence there will
be a subsequence s µ , converging to s ? , where s µ is fixed under the
mapping s → s ( s , λ(µ)).
Now suppose s ? were not an equilibrium point. Then if
s ? = (s?1 , ..., s?n ) some component s?i must be non-optimal against the
others, which means s?i uses some pure strategy πiα which is non-
optimal. [see eq (4), pg. 4] This means that piα ( s ? ) < qi ( s ? ) which
justifies writing piα ( s ? ) − qi ( s ? ) < −.

From continuity, if µ is large enough, |[piα (s µ ) − qi (s µ )] − [piα (s ? ) −

 1  1
qi (s ? )]| < and < . Adding, piα ( s µ ) − qi ( s µ ) + < 0
2 λ(µ) 2 λ(µ)
which is simply φiα ( s µ , λ(µ)) < 0, whence φ+ iα ( s µ , λ(µ)) = 0, whence
c0iα ( s µ , λ(µ)) = 0. From this last equation we know that πiα is not
used in s µ since s µ = πiα c0iα ( s µ , λ(µ)), because s µ is a fixed point.
P
α

And since s µ → s ? , πiα is not used in s ? , which contradicts our

assumption.

Hence s ? is indeed an equilibrium point.

6
4 Symmetries of Games

An automorphism, or symmetry, of a game will be a permutation of

its pure strategies which satisfies certain conditions, given below.
If two strategies belong to a single player they must go into two
strategies belonging to a single player. Thus if φ is the permutation
of the pure strategies it induces a permutation ψ of the players.
Each n-tuple of pure strategies is therefore permuted into another
n-tuple of pure strategies. We may call χ the induced permutation of
these n-tuples. Let ξ denote an n-tuple of pure strategies and pi (ξ)
the pay-off to player i when the n-tuple ξ is employed. We require
that if

j = iΨ then pj (ξ χ ) = pi (ξ)

which completes the definition of a symmetry.

The permutation φ has a unique linear extension to the mixed
strategies.
If si = ciα πiα we define (si )φ = ciα (πiα )φ .
P P
α α
The extension of φ to the mixed strategies clearly generates an
extension of χ to the n-tuples of mixed strategies. We shall also denote
this by χ.
We define a symmetry n-tuple s of a game by

sχ = s ∀χ

it being understood that χ means a permutation derived from a sym-

metry φ.

7
Theorem 4. Any finite game has a symmetric equilibrium point.
P
πiα
α
Proof. First we note that si0 = P has the property (si0 )φ = sj0
1
α
where j = i , so that the n-tuple s 0 = (s10 , s20 , ..., sn0 ) is fixed under
ψ

any χ; hence any game has at least one symmetric n-tuple.

If s = (s1 , ..., sn ) and t = (t1 , ..., tn ) are symmetric then
 
s+t s1 + t1 s n + tn
= , ..., is so too because s χ = s ⇔ sj =
2 2 2
φ φ
 φ
φ s j + t j (s i ) + (ti ) si + t i
si where j = iψ , hence = = , hence
 χ 2 2 2
s+t s+t
= .
2 2
This shows that the set of symmetric n-tuples is a convex subset of
the space of n-tuples since it is obviously closed.
Now observe that for each λ the mapping s → s 0 ( s , λ) used in
the proof of existence theorem was intrinsically defined. Therefore,
if s 2 = s 0 ( s 1 , λ) and χ is an automorphism of the game we will
have s χ2 = s 0 ( s χ1 , λ). If s 1 is symmetric s χ1 = s 1 and therefore
s χ2 = s ( s 1 , λ) = s 2 . Consequently this mapping maps the set of
symmetric n-tuples into itself.
Since this set is a cell there must be a symmetric fixed point s λ .
And, as in the proof of the existence theorem we could obtain a limit
point s ? which would have to be symmetric.

8
5 Solutions

We define here solutions, strong solutions, and sub-solutions. A non-

cooperative game does not always have a solution, but when it does
the solution is unique. Strong solutions are solutions with special
properties. Sub-solutions always exist and have many of the proper-
ties of solutions, but lack uniqueness.

Si will denote a set of mixed strategies of player i and L a set of

n-tuples of mixed strategies.

Solvability:

A game is solvable if its set, L , of equilibrium points satisfies the

condition

( t ; γi ) ∈ L and s ∈ L ⇒ ( s ; γi ) ∈ L ∀i (1)

This is called the interchangeability conditions. The solution of a

solvable game is its set, L , of equilibrium points.

Strong solvability:

A game is strongly solvable if it has a solution, L , such that ∀i

s ∈ L and pi ( s ; γi ) = pi ( s ) ⇒ ( s , γi ) ∈ L

and then L is called a strong solution.

Equilibrium Strategies:

In a solvable game let Si be the set of all mixed strategies si such

9
that for some t the n-tuple ( t ; si ) is an equilibrium point. [si is
the ith component of some equilibrium point.] We call Si the set of
equilibrium strategies of player i.

Sub-solutions:

If L is a subset of the set of equilibrium points of a game and

satisfies condition (1); and if L is maximal relative to this property
then we call L a sub-solution.
For any sub-solution L we define the ith factor set, Si , as the set
of all si ’s such that L contains ( t ; si ) for some t .
Note that a sub-solution, when unique, is a solution; and its factor
sets are the sets of equilibrium strategies.

Theorem 2. A sub-solution,
L , is the set of all n-tuples (s1 , s2 , ..., sn ) such that each si ∈ Si where
Si is the ith factor set of L . Geometrically, L is the product of its
factor sets.

Proof. Consider such an n-tuple (s1 , ..., sn ). By definition ∃ t 1 ,t 2 , ...,

t n such that for each i (t i , si ) ∈ L . Using the condition (1) n-1 times
we obtain successively (t 1 ; s1 ; s2 ) ∈ L , ..., (t 1 ; s1 ; s2 ; s3 ; ...; sn ) ∈ L and
the last is simply (s1 , s2 , ..., sn ) ∈ L , which we need to show.

Theorem 3. The factor sets S1 , S2 , ..., Sn of a sub-solution are closed

and convex as subsets of the mixed strategy spaces.

Proof. It suffices to show two things:

si + s0i
(a) if si and s0i ∈ Si then s?i = ∈ Si ;
2

10
(b) if s# #
i is a limit point of Si then si ∈ Si .

Let t ∈ L .
Then we have pj ( t ; si ) ≥ pj ( t ; si ; γj ) and pj ( t ; si ) ≥ pj ( t ; s0i ; γj )
for any γj , by using the criterion of (1), pg.3 for an eq. pt. Adding
these inequalities, using the linearity of pj (s1 , ..., sn ) ∈ Si , and divid-
si + s0i
ing by 2, we get pj ( t ; si ) ≥ pj ( t ; si ; γj ) since si =
? ? ?
. From this
2
we know that ( t ; s?i ) is an eq. pt. for any t ∈ L . If the set of all
such eq. pts. ( t ; s?i ) is added to L the augmented set clearly satisfies
condition (1), and since L was to be maximal it follows that s?i ∈ Si .
To attack (b) note that the n-tuple ( t ; s#
i ), where t ∈ L will be
a limit point of the set of n-tuples of the form ( t ; si ) where si ∈ Si ,
since s#
i is a limit point of Si . But this set is a set of eq. pts. and
hence any point in its closure is an eq. pt., since the set of all eq.
pts. is closed [ see pg.3 ]. Therefore ( t ; s#
i ) is an eq. pt. and hence
s#
i ∈ Si from the same argument as for si .
?

Values:

Let L be the set of equilibrium points of a game.

We define vi+ = max[pi ( s )], vi− = min[pi ( s )]. If vi+ = vi− we write
s ∈L s ∈L
vi = vi+ = vi− . vi+ is the upper value to player i of the game; vi− the
lower value; and vi the value, if it exists.
Values will obviously have to exist if there is but one equilibrium
point.
One can define associated values for a sub-solution by restricting
L to the eq. pts. in the sub-solution and then using the same defining
equations as above.
A two-person zero-sum game is always solvable in the sense defined
above. The sets of equilibrium strategies S1 and S2 are simply the
sets of "good" strategies. Such a game is not generally strongly solv-

11
able; strong solutions exist only when there is a "saddle point" in pure
strategies.

Simple Examples
These are intended to illustrate the concepts defined in the paper and
display special phenomena which occur in these games.

The first player has the roman letter strategies and the pay-off to
the left, etc.

Ex.1 5 aα −3 Weak
  Solution:
9 7 7 10
−4 aβ 4 a + b, α + β
16 16 17 17
−5 bα 5
−5 +1
3 bβ −4 v1 = , v2 =
17 2

Ex.2 1 aα 1
Strong Solution: (b, β)
−10 aβ 10
v1 = v2 = −1
10 bα −10
−1 bβ −1
Ex.3 1 aα 1 Unsolvable; equilibrium points (a, α), (b, β) and
a b α β
−10 aβ −10 ( + , + ).
2 2 2 2
−10 bα −10 The strategies in the last case have maxi-min and
1 bβ 1 mini-max properties.

Ex.4 1 aα 1
0 aβ 1 Strong Solution: all pairs of mixed strategies.
1 bα 0 v1+ = v2+ = 1, v1− = v2− = 0
0 bβ 0

12
 Ex.5 1 aα 2 Unsolvables
 eq. pts. (a, α), (b, β) and
1 3 3 5
−1 aβ −4 a + b, α + β .
4 4 8 8
−4 bα −1 However, empirical tests show a tendency toward
2 bβ 1 (a, α).

Ex.6 1 aα 1
0 aβ 0 Eq. pts.: (a, α) and (b, β), with (b, β) an example
0 bα 0 of instability.
0 bβ 0

6 Geometrical Form of Solutions

In the two-person zero-sum case it has been shown that the set of
"good" strategies of a player is a convex polyhedral subset of his strat-
egy space. We shall obtain the same result for a player’s set of equi-
librium strategies in any solvable game.

Theorem 5. The sets S1 , S2 , ..., Sn of equilibrium strategies in a solv-

able game are polyhedral convex substs of the respective mixed strategy
spaces.

Proof. An n-tuple s will be an equilibrium point if and only if for every

i
pi ( s ) = max piα ( s ) (1)
α

which is condition (3) on page 4. An equivalent condition is for ever

i and α.
pi ( s ) − piα ( s ) ≥ 0. (2)

Let us now consider the form of the set Sj of equilibrium strategies,

sj , of player j. Let t be any equilibrium point, then ( t ; sj ) will be an
equilibrium point if and only if sj ∈ Sj , from Theo. 2. We now apply
condition (2) to ( t ; sj ), obtaining

13
 sj ∈ Sj ⇐⇒ ∀i, α pi ( t , sj ) − piα ( t , sj ) ≥ 0. (3)

Since pi is n-linear and

t is constant these are a set of linear inequalities of the form Fiα (sj ) ≥
0. Each such inequality is either satisfied for all sj or for those lying
on and to one side of some hyperplane passing through the strategy
simplex. Therefore, the complete set [which is finite] of conditions
will all be satisfied simultaneously on game convex polyhedral subset
of player j’s strategy simplex. [Intersection of half-spaces.]
As a corollary we may conclude that Sk is the convex closure of a
finite set of mixed strategies [vertices].

14
7 Dominance and Contradiction Methods

We say that s0i dominates si if pi ( t ; s0i ) > pi ( t ; si ) for every t .

This amounts to saying that s0i gives player i a higher pay-off than si
no matter what the strategies of the other players are. To see whether
a strategy s0i dominates si it suffices to consider only pure strategies
for the other players because of the n-linearity of pi .

It is obvious from the definitions that:

no equilibrium point can involve a dominated strategy si .

The domination of one mixed strategy by another will always entail

other dominations. For suppose s0i dominates si and ti uses all of the
pure strategies which have a higher coefficient in si than in s0i . Then
for a small enough p > 0

t0i = ti + p(s0i − si )

is a mixed strategy; and t0i dominates ti by linearity.

One can prove a few properties of the set of undominated strategies.

It is simply connected and is formed by the union of some collection
of faces of the strategy simplex.

The information obtained by discovering dominances for one player

may be of relevance to the others, insofar as the elimination of classes
of mixed strategies as possible components of an equilibrium point is
concerned. For the t ’s whose components are all undominated are
all that need be considered ans this eliminating some of the strategies
of one player may make possible the elimination of a new class of
strategies for another player.

15
 Another procedure which may be used in locating equilibrium points
is the contradiction-type analysis. Here one assumes that an equilib-
rium point exists having component strategies lying within certain
regions of the strategy spaces and proceeds to deduce further condi-
tions which must be satisfied if the hypothesis is true. This sort of
reasoning may be carried through several stages to eventually obtain
a contradiction indicating that there is no equilibrium point satisfying
the initial hypothesis.

16
8 A Three-Man Poker Game

As an example of the application of our theory to a more or less

realistic case we include the simplified poker game given below. The
rules are as follows:

(1) The deck is large, with equally many high and low cards, and a
hand consists of one card.

(2) Two chips are used to ante, open, or call.

(3) The players play in rotation and the game ends after all have
passed or after one player has opened and the others have had a
chance to call.

(4) If no one bets the antes are retrieved.

(5) Otherwise the pot is divided equally among the highest hangs
which have bet.

We find it more satisfactory to treat the game in terms of quantities

we call "behavior parameters" than in the normal form of "Theory of
Games and Economic Behavior." In the normal form representation
two mixed strategies of a player may be equivalent in the sense that
each makes the individual choose each available course of action in
each particular situation requiring action on his part with the same
frequency. That is, they represent the same behavior pattern on the
part of the individual.
Behavior parameters give the probabilities of taking each of the
various possible actions in each of the various possible situations which
may arise. Thus they describe behavior patterns.
In terms of behavior parameters the strategies of the players may
be represented as follows, assuming that since there is no point in
passing with a high card at one’s last opportunity to bet that this will

17
not be done. The greek letters are the probabilities of the various acts.

First Moves Second Moves

κ Call III on low
α Open on high
I λ Call II on low
β Open on low
µ Call II & IIIon low
γ Call I on low
ν Call III on low
II δ Open on high
ξ Call III & I on low
 Open on low
ζ Call I & II on low
η Open on low
III Player III never gets a second move.
θ Call I on low
ι Call II on low

We locate all possible equilibrium points by first showing that most

of the greek parameters must vanish. By dominance mainly with a
little contradiction-type analysis β is eliminated and with it go γ, ζ,
and θ by dominance. Then contradictions eliminate µ, , ι, λ, κ, and
ν in that order. This leaves us with α, δ, , and η. Contradiction
analysis shows that none of these can be zero or one and thus we
obtain a system of simultaneous algebraic equations. The equations
happen to have but one solution with the variables in the range (0,1).

We get
√
21 − 321 5α + 1 5 − 2α 4α − 1
α= ,η= ,δ= , and  = .
10 4 5+α α+5

These yield α = 0.308, η = 0.635, δ = 0.826, and  = 0.044.

18
 Since there is only one equilibrium point the game has values; these
−(1 + 17α) −(1 − 2α)
are v1 = −0.147 = , v2 = −0.096 = , and
8(5 + α) 4
 
79 1 − α
v3 = 0.243 = .
40 5 + α

Investigations of the coalition powers yields the following "good

strategies" and values for the various coalitions. Parameters not men-
tioned are zero.

I & II versus III

α = 34 morespaci ι = 14 , 0 ≤ η ≤ 23 space
1
δ =  = 1morespaci value to III : 0.03125 = 32

II & III versus I

δ = 1,  = 0mospace α = 23 morespac
η = 23 mospace value to I : −0.1667 = − 16 sp

I & III versus II

high low
3
··· η =0 11 i δ= 7 3
11 , = 11
bet pass
13 8
pass pass · · · η = 16 11 i 5
value to II : −0.1136 = − 44

The coalition members have the power to agree upon a pattern of

play beefore the game is played. This advantage becomes significant
only in the case of coalition I III where III may open after two passes
when I had planned to pass on both high and low but will not open

19
if I had planned to bet if he got high. The values given are, of couse,
what the single player assures himself with his "safe" strategy.
A more detailed treatment of this game is being prepared for publi-
cation elsewhere. This will consider different relative sizes of ante and
bet.

20
9 Motivation and Interpretation

In this section we shall try to explain the significance of the concepts

introduced in this paper. That is, we shall try to show how equilibrium
points and solutions can be connected with observable phenomena.

The basic requirements for a non-cooperative game is that there

should be no pre-play communication among the players [unless it has
no bearing on the game]. Thus, by implication, there are no coalitions
and no side-payments. Because there is no extra-game utility [pay-off]
transfer, the pay-offs of different players are effectively incomparable;
if we transform the pay-off functions linearly: p0i = ai pi + bi , where
ai > 0 the game will be essentially the same. Note that equilibrium
points are preserved under such transformation.

We shall now take up the "mass-action" interpretation of equilib-

rium points. In this interpretation solutions have no great significance.
It is unnecessary to assume that the participants have full knowledge
of the total structure of the game, or the ability and inclination to go
through any complex reasoning processes. But the participants are
supposed to accumulate empirical information on the relative advan-
tages of the various pure strategies at their disposal.

To be more detailed, we assume that there is a population [in the

sense of statistics] of participants for each position of the game. Let
us also assume that the "average playing" of the game involves n par-
ticipants selected at random from the n populations, and that there is
a stable average frequency with which each pure strategy is employed
by the "average member" of the appropriate population.

Since there is to be no collaboration between individuals playing

21
in different positions of the game, the probability that a particular
n-tuple of pure strategies will be employed in a playing of the game
should be the product of the probabilities indicating the chance of
each of the n pure strategies to be employed in a random playing.

Let the probability that πiα will be employed in a random playing

of the game be ciα , and let si = ciα πiα , s = (s1 , s2 , ..., sn ). Then
P
α
the expected pay-off to an individual playing in the ith position of the
game and employing the pure strategy πiα is pi ( s ; πiα ) = piα ( s ).

Now let us consider what effects the experience of the participants

will produce. To assume, as we did, that they accumulated empirical
evidence on the pure strategies at their disposal is to assume that
those playing in position i learn the numbers piα ( s ). But if they
know these they will employ only optimal pure strategies, i.e, those
pure strategies πiα such that

piα ( s ) = max piβ ( s ).

β

Consequently since si expresses their behavior si attaches positive co-

efficients only to optimal pure strategies, so that

πiα is used in si ⇒ piα ( s ) = max piβ ( s ).

β

But this is simply a condition for s to be an equilibrium point.

[see (4), pg. 4]

Thus the assumptions we made in this "mass-action" interpretation

lead to the conclusion that the mixed strategies representing the av-
erage behavior in each of the populations form an equilibrium point.

The populations need not be large if the assumptions still make

22
it: ... will hold 1 . There are situation in economics or international
politics in which, effectively, a group of interests are involved in a
non-cooperative game without being aware of it; the non-awareness
helping to make the situation truly non-cooperative.
Actually, of course, we can only expect some sort of approximate
equilibrium, since the information, its utilization, and the stability of
the average frequencies will be imperfect.
We now sketch another interpretation, one in which solutions play
a major role, and which is applicable to a game played but once.
We proceed by investigating the question: what would be a "ra-
tional" prediction of the behavior to be expected of rational playing
the game in question? By using the principles that a rational predic-
tion should be unique, that the players should be able to deduce and
make use of it, and that such knowledge on the part of each player
of what to expect the others to do should not lead him to act out of
conformity with the prediction, one is led to the concept of a solution
defined before.
If S1 , S2 , ..., Sn were the sets of equilibrium strategies of a solvable
game, the "rational" prediction should be: "The average behavior of
rational men playing in position i would define a mixed strategy si in
Si if an experiment were carried out."
In this interpretation we need to assume the players know the full
structure of the game in order to be able to deduce the prediction for
themselves. It is quite strongly a rationalistic and idealizing interpre-
tation.
In an unsolvable game it sometimes happens that good heuristic
reasons can be found for narrowing down the set of equilibrium points
1
original text partially erased

23
to those in a single sub-solution, which then plays the role of a solution.

24
 In general a sub-solution may be looked at as a set of mutually
compatible equilibrium points, forming a coherent whole. The sub-
solutions appear to give a natural subdivision of the set of equilibrium
points of a game.

25
10 Applications

The study of n-person games for which the accepted ethics of fair
play imply non-cooperative playing is, of course, an obvious direction
in which to apply this theory. And poker is the most obvious target.
The analysis of a more realistic poker game than our very simple model
should be quite an interesting affair.
The complexity of the mathematical work needed for a complete
investigation increases rather rapidly, however, with increasing com-
plexity of the game; so that it seems that analysis of a game much
more complex than the example given here would only be feasible
using approximate computational methods.
A less obvious type of application is the study of cooperative games.
By a cooperative game we mean a situation involving a set of players,
pure strategies, and pay-offs as usual; but with the assumption that
the players can and will collaborate as they do in the von Neumann
and Morgenstern theory. This means the players may communicate
and form coalitions which will be enforced by an umpire. It is un-
necessarily restrictive, however, to assume any transferability, or even
comparability of the pay-offs [which should be in utility units] to dif-
ferent players. Any desired transferability can be put into the game
itself instead of assuming it possible in the extra-game collaboration.
The writer has developed a "dynamical" approach to the study of co-
operative games based upon reduction to non-cooperative form. One
proceeds by constructing a model of the pre-play negotiation so that
the steps of negotiation become moves in a larger non-cooperative
game [which will have an infinity of pure strategies] describing the
total situation.
This larger game is then treated in terms of the theory of this paper

26
[extended to infinite games] and if values are obtained they are taken
as the values of the cooperative game. Thus the problem analyzing
a cooperative game becomes the problem of obtaining a suitable, and
convincing, non-cooperative model for the negotiation.
The writer has, by such a treatment, obtained values for all finite
two person cooperative games, and some special n-person games.

27
Bibliography

[1] von Neumann, Morgenstern, "Theory of Games and Economic Be-

havior", Princeton University Press, 1944

[2] J. F. Nash, Jr., "Equilibrium Points in N-Person Games", Proc. N.

A. S. 36 (1950) 48-49.

11 Acknowledgements

Drs. Tucker, Gale, and Kuhn gave valuable criticism and suggestions
for improving the exposition of the material in this paper. David
Gale suggested the investigation of symmetric games. The solution of
the Poker model was a joint project undertaken by Lloyd S. Shapley
and the author. Finally, the author was sustained financially by the
Atomic Energy Commission in the period 1949-50 during which this
work was done.

28