Professional Documents
Culture Documents
Addition - Unbounded Rationality
Addition - Unbounded Rationality
North-Holland
A model of decision making under bounded rationality is presented that combines satisficing
behavior with learning and adaptation through environmental feedback. The aspirations, or
goals, of the decision maker dynamically adjust in response to the observed sequence of past
decisions and their corresponding effects on the decison maker’s objective function. A simple
linear response model is employed to represent the beliefs of the decison maker concerning the
causal connection between his/her decisions and the resulting objective function value. The
combination of these simple elements yields a decision process model rich in dynamic behavior;
it can exhibit optimizing behavior in the long-run and chaotic pseudo-random search in the
short-run. As such, the model bridges the gap between substantive rationality and procedural
rationality.
1. Introduction
Correspondmcr TV: Kent D. Wall, DRMI, code 64Wa, U.S. Naval Postgraduate School,
Monterey, CA 93943. USA.
*The author wishes to thank James D. Hamilton and Maxim Engers for their help in framing
the issues presented in the introduction, and in an earlier version of the proof of Theorem I.
Thanks are also due the anonymous referees for many constructive suggestions.
2. Bounded rationality
The descriptive model of decision making presented here owes its con-
ceptualization to Simon’s theory of bounded rationality, the essence of which
may be captured in eight statements:
x(t+1)=x(t)+dx(t+l)=x(t)+cc(t+l)d(t+l), (I)
where d(t) represents the local search direction and a(t) denotes the step
length to be taken along this direction. Furthermore, these two quantities are
chosen on the basis of an information set Y(t) gathered from experience:
ci(t + 1) = Yv,(.F(t)).
334 K.D. Wall, A model ofdecision making
Realistic modeling would suggest that Y(t) be very simple and limited in its
scope. The decision maker should not be assumed in possession of inform-
ation not directly observable or easily inferred from directly observable data.
For example, in a model of monopolistic producer behavior we would
exclude elasticities of demand from Y(t), but include the past prices asked
for the product produced and the corresponding consumption observed.
In consideration of [D] through [F] the decision maker is presumed to
hold at the end of time t an aspiration level a(t+ 1) representing a value of
the objective function f( .) which he or she believes is satisfactory and a
reasonable goal to attain in the next time period. Motivation for continued
search is provided by failure to achieve the currently defined goal; i.e.,
f’(t+l)=f(t)+c:dx(t+l). (8)
Mr+j)=C4r+ l)-f(mlllct~~2
The step length is seen to depend upon the difference between the achieved
objective and the desired value of the objective. Thus, if the decision maker is
close to his or her goal then a small step is taken, while large discrepancies
between the actual and desired objective value will lead to more bold search,
in the form of large changes in the decision variables. If the decision maker
had achieved the goal then no step is taken and satisticing behavior obtains.
The search direction is seen to be a unit vector which approximates the
direction of steepest ascent but uses only ‘backward looking’ information.
336 K.D. Wall, A model oidecision making
While this unit vector never changes in magnitude, the direction in which it
points does change with experience and therefore embodies the learning and
adaptation required by [G].
The N-vector c, represents the decision maker’s best estimate of the
sensitivity of f(t) to changes in x(t). It is comprised of nothing more than
simple finite first differences derived from F(t). This formulation follows after
the work of Cyert and March (1962) and Day (1967) who argue that humans
employ finite first difference approximations when dealing with the concepts
of derivative and gradient. In the one-dimensional case where x(t) is a scalar
c, is simply
c,=c(t)=[f(t)-f(t-l)]/[x(t)-x(t-l)].
Af(t-N+l)=c;[x(t-N+l)--(t-N)],
A f(t) = X(t)c,,
c,=X(t)+Af(t).
If the rows of X(t) are linearly independent when X(t)’ is replaced by X(t)-‘.
All that remains to be specified is the exact form of (6), and here we appeal
to March and Simon (1958) and March (1988) as a point of departure.
Aspirations are hypothesized to adjust to experience by a linear first-order
adaptive adjustment mechanism. Two additional terms, however, are added
representing environmental feedback from the observed values of f(t) and
the observed rate of change in f(t). The result is a form of double
exponential smoothing on f(t):
~(t+l)=Cl-ljlP(t)+Bf(t),
i(t+l)=[l-s]i(t)+s[f(t)-f(t-l)],
a(t+1)=[1-~]a(t)+~f(t+1)+~i(t+1).
but also the perceived rate of change of the objective function that influence
aspirations. Goal formation is argued to be significantly affected by the
magnitude of change of the objective with a rapidly rising (falling) f(t)
leading the decision maker to increase (decrease) the value of a(t+ 1) above
(below) what would have been the case by keying only on the level of f(t).
The inclusion of a rate of change effect does not appear to have been
considered in the existing literature, but is crucial to the performance of this
model. The motivation for this term is based on two considerations. First,
humans attend to rapidly changing factors far more readily than they do to
those changing only gradually over time [Kahneman and Tversky, in
Kahneman, Slavic and Tversky (1992, ch. 4), and Hogarth (1987, ch. 2)]. One
needs only to reflect upon the public reaction to a change in gasoline prices
from, say $0.90/gal to $1.30/gal that occurs over a period of two or three
weeks. The same magnitude of change that occurs over a period of two years
evokes almost no reaction at all. Second the formation of goals is more
dependent upon long term perceptions than the formation of f’, which is
only concerned with what can be expected over the next decision period.
Therefore, aspirations are more likely to be a function of what is perceived to
be possible to achieve over more than one decision period. Using f(t+ 1) as
a base and i(t+ 1) as measure of the average rate of change in f(t) over the
next k periods, one may expect that it is possible to achieve l(t+ l)+i(t+ 1)k
over the next k periods. By setting y =+k we obtain the expression for
a(t + 1).
The model of decision making under bounded rationality can now be
summarized:
Step 0. Initialization.
Step I. Update the information set at end of period t (by observing f(t)
and remembering x(t))
u(t+l)=[l-~]a(t)+qqt+l)+yi(t+l). (12)
338 K.D. Wall, A model qf decision making
c,=X(t)‘df(t). (13)
(14)
(15)
h(t+l)=a(t+l)d(t+l), (16)
x(t+l)=x(t)+dx(t+l). (17)
This function has a unique global maximum at x* = 4.0 with f * = 15.0. The
340 K.D. Wall, A model of decision making
J
0 10 20 30 40 50
Fig. 1. Aspirations, expectations, and profit as functions of time for the quadratic function in
one-dimension.
42) X”‘X
f(C)+ h?(C) v---v
f(t)
Fig. 2. The decision variable as a function of time for the quadratic problem in one-dimension.
342 K.D. Wall, A model of decision making
’ 0 10 20 30 40 50
Time Index
Fig. 3. Aspirations, expectations, and profit as functions of time for the quadratic problem with
inadequate learning leading to suboptimal decisions.
a(t) X”‘X
j‘(t) + hi(t) v---v
f(t)
m
0 10 20 30 40 50
Time Index
Fig. 4. The decision variable as a function of time with inadequate learning and adaptation
leading to suboptimal satisficing decisions.
K.D. Wall, A model of decision making 343
Time Index
Fig. 5. Aspirations, expectations, and achieved profit for the quadratic problem with time
varying environment.
4) X”‘X
j‘(C)+ k(t) v---v
f(t)
0
0 20 40 60 60 100 120 140 160 160 200
Time Index
Fig. 6. The decision variable as a function of time for the quadratic problem with time varying
environment.
JEB.0 C
K.D. Wall, A model of decision making
-~02C(XZ(t)w1Xl(t))/00182
- 5009
and adds this to its information set composed of past observations on f(t)
and x(t). It then decides on a new value for x and implements this for period
t + 1. The firm is not assumed to have any information concerning the form
or parameters of the demand function and is not assumed to know its supply
function.
The difficulty presented to our model of decision making can be gauged
from the contour plot of f(x) displayed in tig. 7. Profit falls off rapidly for
K.D. Wall, A model of decision making 345
capital-to-labor ratios less than 0.15 and production rates less than 15,000
units. The ‘ridges’ encountered just prior to these two precipices often
confound many search algorithms by inducing zig-zagging behavior. More
importantly, the protit function is strictly concave only within a neighborhood
of x* = CO.3067; 47,274.41]’ (at which point f* = $183,662.7). Thus for large
values of x1 and x2, strict concavity does not hold.
The first experiment employs an initial aspiration level of a(O)=$150,000
and initializing decisions of
The parameters for the adjustment eqs. (10)+12) are set at:
Fig. 8 displays the decision history in terms of x,(t) versus x,(t). The
model finds the neoclassical optimizing solution after approximately 150
iterations and, for all practical purposes, converges. The value of the profit
function remains within one decimal place of the optimum, x1 is within two
decimal places of its optimizing value, and x2 matches its optimizing value to
within 0.50/,.
As with the first simulation example, one might ask what the behavior of
the model is if it is embedded in a changing environment. Would it ever be
able to find a time-varying optimum? Would the decision variables display
behavior that could be interpreted as a sample from some stochastic process?
Some food for thought is provided by figs. 9 and 10. Here the model
operates in a shifting environment much like that leading to figs. 5 and 6.
Two shifts in the parameters of the profit function are introduced, corres-
ponding to shifts in the demand curve and in the factor input supply curves.
One shift occurs at t = 120 and the other at t =250. Over the period of the
simulation there are now three different maxima that must be sought:
Capital-Labor Ratio
Fig. 8. Convergent behavior in decision space for the two-dimensional profit maximization
problem.
For this simulation experiment all model parameters are kept at their
previous values except C$= 0.05.
The model locates the three maxima to within some small neighborhood.
Fig, 9 presents the path taken in decision variable space, clearly illustrating
how search proceeds from one solution to the next as information is received
discontirming prior conceptions of where the optimal decision is located. Fig.
10 shows that the model requires approximately 100 time periods to learn
where the optimal decision resides.
5. Estimation framework
For the model to be truly useful in empirical research it must permit
estimation and testing in conjunction with time series data. This requires
casting eqs. (9)-(17) in a form suitable for econometric work and one
framework immediately suggests itself. Eqs. (lo)-( 12) comprise a recursive
K.D. Wall, A model of decision making 341
Capital-Labor Ratlo
Fig. 9. Behavior in decision space for the two-dimensional problem with time varying
environment.
In
“0 * f’ ” ” ”
x I
nme index
Fig. 10. Aspirations, expectations, and achieved prolit as functions of time for the two-
dimensional profit maximization problem with time varying environment.
a(t) X”‘X
f(t) + hi(t) v---v
f(t)
JE.90. D
348 K.D. Wall, A model of decision making
F=
r(1-B)0 0 (l-6)
01
0 ,G=
ra 6 -6
0 o...o
1
O...O ,
1 4
H(z(t))=CO:O:ctlllCt1121,
Y (l-4)
1 1
0 0 o...o
1
and y(t) =x(t). The state space model is linear in the state but time-varying in
the output equations because of the terms involving c,/ll c, 11’. It should be
noted, however, that this time variation in the coefficient matrix H is due
solely to exogenous variables and may be treated as predetermined. Estimat-
ing the model can be accomplished by a number of techniques, for example
that used in Burmeister, Wall and Hamilton (1986), Kalaba and Tesfatsion
(1980, 1988), or Pagan (1980). If one is willing to state stochastic hypotheses
regarding additive disturbances in (18)419), then a Gaussian likelihood
function can be postulated and the methods of the first and last references
above apply. If one does not wish to introduce additive random errors then
a least-squares approach can be taken and the methods of Kalaba and
Tesfatsion applied. In either case, the methods are well developed, tried and
true.
The researcher applying this estimation framework needs to obtain time
series data for f(t) and x(t) over the period of interest, say, {to ZL~:~}, and
use them to form the z(t) vector. The c, are constructed using (13) and the
first N observations of f(t) and x(t) as initial conditions. The required
pseudo-inverses can be accomplished by singular value decomposition.
Finally, H(z(t)) and h(z(t)) are formed. The need for initial conditions in
K.D. Wall, A model cfdecision making 349
computing the c, means estimation proceeds using the data over the interval
I lo + N - 1 stst/ }. Four parameters are estimated: {#,/I, 6,~). In addition,
.~n estimate of the trajectory of a(t) over the sample period can be obtained
by coupling the parameter estimation algorithm to a state space mode1 filter/
smoother algorithm like that of De Jong (1989). This allows the all
important goal formation behavior to be revealed and can be used to shed
light on how aspirations have responded to stimuli over the sample.
As an illustration of formulating an estimation problem, suppose one
wishes to study the behavior of the U.S. auto industry over the past four
decades, with particular interest in how the industry responded to growing
foreign competition. In this case behavior might be interpreted in terms of
the decisions taken to adjust production capacity, employment, and price.
Thus x(t) is a 3 x 1 vector composed of time series observations on each of
thcsc three variables. Furthermore, it is hypothesized that decisions are based
upon attention to profits; thus S(t) is the observed industry profit in period t.
If it is hypothesized that decision makers focus on market share, then f(t) is
the observed market share of domestic auto producers. Depending on the
frequency of observation, t is indexed by months, quarters, or years.
l.\timation of this mode1 requires four time series, and then their use in
generating the three time series comprising the elements of c,. The four
ehtimated parameters, together with the estimated trajectory for a(t), would
yield information on the rate with which the industry responded to the
foreign challenge, how industry goals were affected by changing market
conditions, or whether profit or market share better reflects the concerns of
industry management.
Appendix
x(~)-x(O)=C.~*--~‘(X(O))IC~IIICO
112=Df’(~*)C~*-x(0)lco/llco II*>
and
11x(l)-40) //
where P = 11
DJ‘(g*) ((//IDf’(tzz)
I/. But from the strict concavity of f, p < 1 and
so the step to x(l) is not ‘too’ large so as to over
kb’ ;;x:;ic; “i;;(;;;i”)l’; (o))
x .
Now assume that for some t=n, ,f(n)>J‘(n- l)>f’(n-2)>f(n-3)~ ..‘>
f(n- N). By exactly the same arguments as above, ck[x*-x(n)] >O, so
d(n+ 1) is a direction with a component along x*-x(n) and ‘towards’ x*.
Furthermore, I/x(n+l)-x(n))I<IIx*-x(n)11 so that f(n+l)>f(n). Thus f(n)
converges monotonically to some upper bound, F, and by continuity, x(n)
converges to some point z such that F =f’(z). To see that this upper bound is
indeed f‘*, and that z=x*, consider (16) rewritten as
IIc,//C~(~+1)-x(~~l=C.f*-f‘(~)lc,lllc,//
In the limit as PZ-+CG,c,-Df’(d, the lefthand side +O, and c,/IIc,/(+l. This
implies ,f(x(n))-.f* and continuity of f implies x(n)-+z=x*. Q.E.D.
352 K.D. Wall, A model of decision making
References
Baumol, W.J. and R.E. Quandt, 1964, Rules of thumb and optimally imperfect decisions,
American Economic Review 54, 23-46.
Burmeister, E., K.D. Wall and J.D. Hamilton, 1986, Estimation of unobserved expected monthly
inflation using Kalman filtering, Journal of Business and Economic Statistics 4, 147-160.
Crain, W. Mark, William F. Shughart II and Robert D. Tollison, 1984, Journal of Economic
Behavior and Organization 5, 3755386.
Cyert, R.M. and J.G. March, 1962, A behavioral theory of the firm (Prentice-Hall, Englewood
Cliffs, NJ).
Dawes, R. and B. Corrigan, 1974, Linear models in decision making Psychological Bulletin 8 I.
95-106.
Day, R.H., 1967, Profits, learning and the convergence of satisticing to marginalism, Quarterly
Journal of Economics 81, 302-311.
Day, R.H. and E.H. Tinney, 1968, How to cooperate in business without really trying, Journal of
Political Economy 76, 8533600.
De Jong, Piet, 1989, Smoothing and interpolation with the state-space model, Journal of the
American Statistical Association 84, 108551088.
Edwards, Ward, 1968, Conservatism in human information processing, in: B. Kleinmuntz. ed.,
Formal representation of human judgement (Wiley, New York) 17-52.
Einhorn, H., 1980, Learning from experience and suboptimal rules in decision making, in:
T.S. Wallsten, ed., Cognitive processes in choice and decision behavior (Lawrence Erlbaum,
Hillsdale, NJ).
Friedman, Milton, 1953, Essays in positive economics (University of Chicago Press, Chicago).
Hogarth, Robin, 1987, Jugment and choice, 2nd ed. (Wiley, New York).
Kahnemann, Daniel, Paul Slavic and Amos Tversky, eds., 1982, Judgement under uncertainty:
Heuristics and biases (Cambridge University Press, Cambridge, MA).
Kalaba, Robert and Leigh Tesfatsion, 1980, A least-squares model specification test for a class of
dynamic nonlinear economicmodels with systematically varying parameters, Journal of
Optimization Theory and Applications 32, 5388567.
Kalaba, Robert and Leigh Tesfatsion, 1988, Exact sequential filtering, smoothing and prediction
for nonlinear systems, Nonlinear Analysis: Theory, Methods and Applications 12, 5999615.
Levinthal, Daniel and James G. March, 1981, A model of adaptive organizational search,
Journal of Economic Behavior and Organization 2, 307-334.
March, James G., 1988, Variable risk preference and adaptive aspirations, Journal of Economic
Behavior and Organization 9, 5-24.
March, James G. and Herbert A. Simon, 1958, Organizations (Wiley & Sons, New York).
Nelson, Richard R. and S. Winter, 1973, Toward an evolutionary theory of economic
capabilities, American Economic Review 63, 440-449.
Pagan, Adrian, 1980, Some identification and estimation results for regression models with
stochastically varying coefftcients, Journal of Econometrics 13, 341-363.
Radner, Roy, 1975, A behavioral model of cost reduction, Bell Journal of Economics 6, 196-215.
Shoemaker, Paul, 1980, Experiments on decisions under risk: The expected utility hypothesis
(Martinus Niyhoff Publishing, Boston, MA).
Simon, Herbert A., 1955, A behavioral model of rational choice, Quarterly Journal of Economics
69, 99-118.
Simon, H.A., 1957, Models of man (Wiley & Sons, New York).
Winter, S. 1971, Satisficing, selection, and the innovating remnant, Quarterly Journal of
Economics 85,2377261.
Witt, Ulrich, 1986, Firms’ market behavior under imperfect information and economic natural
selection, Journal of Economic Behavior and Organization 7, 265-290.
Zangwill, W.I., 1969, Nonlinear programming: A unified approach (Prentice-Hall, Englewood
Cliffs, NJ).