Sie sind auf Seite 1von 9

CHAPTER 6

Robert Kozelka

A Bayesian Approach
to Jamaican Fishing

IN Davenport's classic article ( 1960), he uses


game theory to analyze strategies in a Jamaican fishing village. Allow­
ing some leeway for the estimated numerical data, his results are in
good accord with an overall minimax strategy by the fishenrten. That
is, he finds that to make the minimum gain as large as possible a cer­
tain mixed strategy is appropriate and that, in fact, the percentage of
fishing captains following each of the pure strategies in the mixed
strategy is as close as possible to this optimal strategy, given the dis­
creteness involved.
The purpose of this paper is to examine further into the idea of
optimal strategies, using, in particular, decision-theoretic notions
rather than game-theoretic ones. The results cannot be any closer to
observation than Davenport's (in fact, they are quite different), but
the assumptions back of the ideas to be presented seem more realistic.
In any case, the various decision-theory concepts may be more use­
ful to the anthropologist than the basic notions of game theory.
A central assumption of two-person zero-sum game theory is that
the second (minimizing) player is both rational and operating in such
a way as to diametrically oppose the interests of the first player. The
definition of rationality is clearly set forth in von Neumann and
Morgenstern ( 1955:28): A rational man operates so as to maximize
expected utility, or, in the game theory situation, to maximize min­
imum expected utility.
A great deal has been written about games against nature; see, for
example, Savage ( 1954) and Blackwell and Girshick ( 1954). It is
very doubtful whether nature consciously maximizes her expected

I wish to express my appreciation to the Summer Seminar in Quantitative


Anthropology for the suggestion that the analysis presented here might be of
interest.
117
118 ROBERT KOZELKA

utility, even assuming nature has a utility function. If we assume that


the first player's utility reflects that of the second (i.e., that nature's
utility for being fished can be measured by expected income in
pounds), it is still generally agreed that any assumption regarding
nature's malevolence is untenable. One should not, therefore, play a
game against "inanimate nature" as one would play against a "ra­
tional" player. How, then, should one play?
Davenport writes, "Let it be quite clear that the writer in no way
professes to understand the mathematics of proof of the theory of
games as set forth by its authors." Similarly the present author for­
goes any quarrel with Davenport's figures. The point of this article
is: Given the estimates of average yield per fishing month, is there a
"better" way to choose an optimal strategy? Although ordinary statis­
tical decision theory deals with losses from certain strategies, the
formulation here is in terms of gains ; one may convert to the stand­
ard statistical format by replacing each gain g1 (average yield), by the
quantity maxgain gi. This in no way changes the appropriate strat­
-

egies under various rules.


Davenport also writes, "There is absolutely no way for the fisher-
men to predict when the current will come or go, nor can they tell
from shore if the current is running or not." If this is the case, one
can think of current as appearing randomly with probability (to use
his figure) 0. 25. It is time to introduce some terminology. We assume
there are two "states of nature": 8 i = current, 8 " = no current. The
fishermen have three "actions": I = fish inside, I-0 = fish in-out,
0 = fish outside. With the states of nature are associated "a priori
probabilities": w' = Pr(e') = 1A, w" = Pr(e") = % . Using Davenport'
figures, one also has a "gain table" or set of expected (monthly) payom
as follows:

1-0 0
8'
8"
1
� ::
1 ____._�_
,___ _ _
1 ·�:�
_�--'---�
���
The reader will recognize the latter simply as Davenport's game
matrix.
Now given no information about the actual state of nature except
the a priori probabilities, the fishermen can calculate their expected
119 A Bayesian Approach to Jamaican Fish i ng

gain, E, from each action using these a priori probabilities. We find .


E[l] = i(l 7 .3) + £(11.5) = 12.95
E[l-0] = 14.05
E[O] = 14.35.
Thus if the fisherman are "rational," in the sense of acting to maximize
expected gain, they should take the one action they never take : Put all
of their pots outside. If, in fact, "both figures for the outside strategy
may be too low" (Davenport's observation), the expected gain would
be even larger for action 0. 1
It is not difficult to see that if rational behavior is maximizing ex­
pected utility, then having only a priori probabilities and a gain table
will always make the optimal strategy a single action, unless two or
more actions lead to the same expected gain. One should note, fur­
thermore, that the expected gain under this optimal strategy is higher
(14.35) than that ( 13.3) achieved by the optimal mixed strategy
derived from game-theoretic considerations.
Why this discrepancy in results? The answer is in the second assump­
tion-the malevolence of nature. If the fishermen always take
action 0 against a malevolent player (nature), that player will always
play action C (current running) and the fishermen will lose 4.4. But
suppose nature is not malevolent but only indifferent. Suppose fur­
ther that nature "tips her hand" in the sense that there is a relation
between current running and observed sea conditions. (We now go
beyond Davenport's data in order to illustrate some further decision­
theoretic concepts.) If e' and()/' continue to represent the states of
nature, let Z' and Z" represent the observations "sea conditions
suggest current running" and "sea conditions suggest current not
running," respectively. Further, suppose that what conditions sug­
gest and the actual occurrence or nonoccurrence of current are com­
patible only Y3 of the time. That is
Pr (Z' I 8') = Pr (Z" I 8") = t, Pr (Z" I 8') = Pr (Z' I 8") =
i
(The numbers are for illustration only, to keep the computation
simple.) On the basis of the sea conditions we can define nine "pure
strategies," which are pairs of actions specifying what to do depending
on whether Z' or Z" is observed. They can be conveniently tabulated
as in Table 1.
120 ROBERT KOZELKA

TABLE 1. PURE STRATEGIES

Strategy No. Actions G' = Exp. Gain \ E>1 G" = Exp. Gain \ E>"
S1 I, I 1 7.3 1 1. 5
S2 I, 1-0 9.23 13.3
Sa I, 0 2.83 14.53
s, 1-0, I 13.26 15.16
s, 1-0, 1-0 5.2 1 7.0
Sa 1-0, 0 -1.2 18.2
s, 0, I 1 0.06 17.56
s. 0, 1-0 2.0 19.4
So 0,0 -4 .4 20.6

In the strategy pairs, the first action entry refers to the action taken
if Z' is observed and the second action entry refers to that taken if
Z" is observed. The expected gain is computed by multiplying the
• probability of taking the specific action by the gain of that action,
given the appropriate state of nature, and then summing over the two
actions. For example, given strategy S3 and state of nature (J ' , one
observes Z' -and hence takes action I -with probability 1/3 and
gains 17 .3, whereas if one observes Z" he takes action 0 -
with probability %-and gains - 4.4. The expected gain is thus
l/3 ( 17 .3) + % ( - 4.4) = 2.8333 . . . = 2.83

The computation for S3 and (J" is 1/3 ( 1 1.5) + % (20.6) 14.53. The =

other values are computed similarly.


It is clear by comparing the entries that the results of some strategies
render those strategies undesirable. If the expected gains for a given
strategy, S;, are smaller for each state of nature than those for strat­
egy S;, we say S., dominates Sror that Sj is not admissible. From
Table 1, Si and S3 are dominated b y S 4, S5 is dominated by S1,
and S6 by S 7 and S8• Hence the admissible pure strategies are
S i, S 4, S1, S8, S9• Of these, S 4 has its smaller gain larger than
the smaller gain of any of the others, so it is the maximin pure
strategy. (If we were dealing with losses, the corresponding idea would
be to minimize the maximum loss, hence a minimax strategy.) Note
that we have not defined the maximin strategy to be optimal, nor
have we required a malevolent opponent. Several different definitions
of optimality have been suggested, of which maximin is only one; (see,
1 21 A Bayesian Approach to Jamaican Fish i ng

for example, Milnor 1954). In particular, maximin does not take ac­
count of the a priori probabilities w' and w" .
For only two possible states of nature, it is instructive and conven­
ient to think of the expected gains as points in two-dimensional space.
These are plotted in Fig. 1. Geometrically, a mixed strategy must
G"

M = (1 4.25, 14.25)

12

FIGURE 1

produce a point on the line segment joining its pure-strategy com­


ponents, so the set of all expected gain points forms a "polyhedral
convex set"; this set is the region bounded by the line segments in the
figure. The cross-hatched line segments indicate the admissible mixed
(and pure) strategies. It is clear from the figure that a strategy is
dominated if it lies below and to the left of another strategy, and that a
pure strategy may be dominated by a mixed 1strategy without being
dominated by any other pure strategy. (For an extensive explanation of
the geometry of the two-state problem, see Chernoff and Moses
[ 1959) .) Furthermore, the maximin strategy can now be found as the
intersection of the line G'= G" with the set of admissible strategies,
and the proportions of the resulting line segment give the proper mix­
ture of these strategies. For these data, the intersection is at the point
1 22 ROBERT KOZELKA

( 14.25, 14.25) and the corresponding mixed strategy is at (0.245Si ,


0.75584). This is indicated in the figure by M. A suitable adjustment
handles the case where the line G' = G" does not intersect the convex
set of expected gain points.
We emphasize three points of the above argument: ( 1) Nowhere
have we defined the maximin strategy to be "optimal." (2) While the
concept of a mixed (randomized) strategy may have some validity for
activities against rational opponents, as in determining hunting grounds
by casting bones, the executive found by his board of directors to be
making decisions by use of a random number table may soon be look­
ing for another job. (3) The a priori probabilities of the states of
nature have not been used.
Returning to von Neumann's definition of rationality, we define an
optimal strategy as one which maximizes "expected'' gain. Actually,
since the expected gain table was constructed, as the name implied,
from expected values, some other expectation must be involved. In
fact, we will consider the expectation by the a priori probabilities of
the entries in the expected gain table. Thus we are really trying to
maximize expected expected gain-call this Bayesian expected gain,
following the statistical terminology. For strategies S1 and S9 (and S5 ,
which is dominated), the Bayesian expected gains are simply the
values computed at the beginning of this paper, since the actions in­
volved are independent of the observations Z' and Z" . The Bayesian
expected gains for all admissible pure strategies are as follows:
Strategy S1

Bayesian expected gain 12. 95 14.69 15.70 15.05 14. 35

These are computed, naturally, by taking the weighted average of the


expected gains G' and G" by the a priori probabilities w' and w"
respectively. Now since the expectation of a weighted sum is that sum
of the component expectations, the Bayesian expected gain of a mixed
strategy must lie between the Bayesian expected gains of the compon­
ents. Thus the maximum Bayesian expected gain (called the Bayes
expected gain corresponding to w' and w" ) must occur at a pure
strategy point (called the Bayes strategy corresponding to w' and w").
In the example, the Bayes expected gain is 15.70, occurring for the
Bayes strategy S1.
1 23 A Bayesian Approach to Jamaican Fishi ng

Again the geometry is instructive. The Bayesian expected gain,


for a given strategy Si, is w' G' + w" G" . For given w' and w" ,
'
w G' + w G"
"
k defines a family of parallel straight lines. These lines
=

have nonpositive slope (since both w ' and w" are between 0 and 1), and
k increases as the family is traversed upward and to the right. Hence
to maximize the Bayesian expected gain by some strategy in the con-
vex set is to find the line w ' G' + w " G" = k ma(J} which is farthest
to the "northeast" and intersects the convex set. This is indicated in Fig.
2. The value of k ma(J} is, of course, 1 5.70, occurring at the strategy
point S7 on the line 1/,i G' + % G" = 15.70. We remark that

G"

........ ..........
12 '•,_, _ _

,
__ __
_
,
_,

'
--
--
-

8 - •• , _
,
_ % G' + 3/4 G" = 9

I � G'
-4 4 8 12 16

FIGURE 2

the maximizing straight line must intersect the strategy set in at


least one pure strategy point, which means that if one makes decisions
on the basis of Bayes strategies one need never consider mixed strat­
egies. Whether or not such a result is desirable or realistic depends on
one's definition of optimality.
1 24 ROBERT KOZELKA

SUMMARY

The purpose of this paper has been to investigate possible defini­


tions of optimality in situations where the underlying assumptions of
game theory seem to be inappropriate. Statisticians have long con­
sidered the idea of hostile nature, required for game theoretic
optimality, to be meaningless or unnecessary. Without some further
probabilistic assumptions, this does away with the concept of a min­
imax strategy in problems like those considered by Davenport in his
paper on a Jamaican fishing village. Such problems, known to the
statistical trade as "no-data" problems, are usually solved by directly
maximizing expected gain (minimizing expected loss). However, the
minimax optimal strategy reappears when decisions take into account
expected gains (losses) based on nondeterministic observations which
are related to nature's "strategies" (states). However, such optimality
leaves out of account any a priori information available. Taking
expectations over these a priori probabilities incorporates all of the
various kinds of rational information. This, together with the usual
optimality definition, results in one (or more) optimal pure strategies
called Bayes strategies with corresponding maximized expected gain.
These ideas have been illustrated geometrically in the two-dimen­
sional problem of Davenport. The results generalize in a straightfor­
ward fashion to problems of higher dimension. For these generaliza­
tions and for the relationship between Bayesian considerations and
classical statistical test procedures the reader is invited to consult
either Chernoff and Moses ( 1959) or Blackwell and Girshick (1954),
depending on the level of his mathematical sophistication.

NOTES
1. It is not clear whether by "too low" Davenport means that the values 20.6 and
- 4.4 should both be replaced by numbers greater than those numbers or by num­
bers farther away from zero. In the former case, the expected gain from strategy 0
would be even larger; in the latter, strategy 0 might no longer be optimal if - 4.4
is replaced by a loss which is larger negatively.
1 25 A Bayesian Approach to Jamaican Fishing

REFERENCES
Blackwell, D., and M. A. Girschick. 1 9 54. Theory of Games and Statistical
Decisions. New York: Wiley.
Chernoff, H. and L. E. Moses. 1 9 59, Elementary Decision Theory, New York:
Wiley.
Davenport, W. 1 960. Jamaican Fishing: A Game Theory Analysis in Papers on
Caribbean Anthropology: Yale U. Publications in Anthropology No. 59, pp.
3-1 1 .
Luce, R . D . , and H . Raiffa. 1 957. Games and Decisions. New York: Wiley.
Milnor, J. 1 954. Games Against Nature in Decision Processes, ed. R. M. Thrall et
al., pp. 49-59. New York: Wiley.
Savage, L. J. 1 954. The Foundations of Statistics. New York: Wiley.
von Neumann, J., and 0. Morgenstern, 1 955. Theory of Games and Economic
Behavior, 3rd. ed. Princeton: Princeton U. Press.

Das könnte Ihnen auch gefallen