Journal of Economic Behavior & Organization 67 (2008) 78–89

Bounded rationality, cognitive maps, and trial and error learning

Abstract
The term “bounded rationality” is used here to connote the reasoning capabilities of an actor who, on the one hand, has a goal
to achieve and an at least partially formed theory as to how to achieve it, and on the other hand, that the theory is somewhat crude,
likely will be revised in the course of the effort, and that success is far from assured. This article presents a theory of how trial and
error learning interact with theory modification in the course of problem solving under bounded rationality. The empirical focus is
on efforts to advance a technology, especially medical practice, but the analysis is quite general. A central question explored is what
makes progress in a field hard or easy.
© 2007 Elsevier B.V. All rights reserved.

JEL classification: O30; O31; O33

Keywords: Bounded rationality; Search; Progress

1. Introduction

In this paper, I use the term “bounded rationality” to connote the reasoning and learning abilities of an actor who
has a goal to achieve and, on the one hand, an at least partially formed theory about how to achieve it (this is the
“rationality” part of the concept), and on the other hand, that the actor’s theory is likely somewhat crude and perhaps
even a bad guide for action, and that success is far from assured (this is the meaning of the “bounded” qualification
to rationality). In the analysis that follows I present a theory of the dynamics of problem solving under this kind of
bounded rationality.
A theory of human and organizational problem solving, which sees the actor as guided by a cognitive map or theory,
is germane to certain kinds of contexts, but not others. Much of the history of cognitive science over the last 30 years has
involved the efforts of many scholars in the field to get away from the kind of “good old fashioned artificial intelligence”
(the term was coined by Dreyfuss and Dreyfuss, 1986) developed by Newell and Simon (1972) and kindred theorists,
which did see problem solving this way (see Nelson and Nelson, 2002 for a capsule history). Virtually all scholars in
the field now would agree that cognitive maps or conscious theories are not needed as part of the explanation for what
humans (or organizations) do in many contexts (see for example Hendriks-Jansen, 1966 or Clark, 1997). However, it
would seem clear that there are important arenas of individual and organizational behavior where mental models or
theories do play an important role (Johnson-Laird, 1983; Holyoke and Spellman, 1993).

E-mail address: rrn2@columbia.edu.

0167-2681/$ – see front matter © 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jebo.2007.06.002
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 79

The presumption that human problem solving proceeds under the guidance of cognitive maps or theories shows up
in several places in social science research. It is central in recent writings about the role of business strategies in guiding
what firms do, and sometimes leading them into dead ends (two apt examples are Leonard-Barton, 1992 and Tripsis
and Gavetti, 2000). Many recent analyses of how engineers solve problems and design things fit this mold (Vincenti,
1990 is wonderful on this). Dosi (1982) and Nelson and Winter (1982) use the terms technological paradigms and
technological regimes to denote the cognitive orientation of technologists working in a field in a particular era. This
view of human problem solving also is prominent in analyses of how scientists deal with problems (see for example
Thagard, 1999). Thomas Kuhn’s theory of the behavior of scientific communities (1962) is concerned both with the
role of prevailing theories (paradigms) in guiding the action of individual scientists, and with the processes by which
theories change.
My focus in this essay is on individual and organizational problem solving in these kinds of contexts. While I believe
that the perspective I develop can provide quite general insight relevant to analysis of problem oriented search guided
by a cognitive map, the details of the theorizing are tailored to suit the empirical arena where I have done most of my
own work, efforts to advance the performance of a technology, and in particular my current focus of that research,
medical innovation.
In the analysis that follows, at any time search for a satisfactory or a better way of doing something is strongly
oriented, but also limited, by a theory held by the actor at that time. Empirical exploration of alternatives at any time
is treated as like going down a path, which current theory suggests is promising, finding out where it in fact leads, and
then perhaps trying another path. However, the actor is not locked into his present theory. As a result of what is learned
in exploration, theory may be revised. In turn this may lead to choice of different kinds of paths. What our explorer is
able to achieve ultimately depends on the efficacy of this interactive learning process, and on luck.
Nothing in this broad theory of learning presumes that our actor in fact does learn effectively. There is no presumption
that the problem solver assesses the empirical evidence accurately or draws a helpful theoretical inference. In their
analyses of business behavior, Leonard-Barton (1992) and Tripsis and Gavetti (2000) are concerned with explaining
why the strategies of the businesses they studied did not respond adequately to what should have been obvious
empirical clues that things were not going well. The history of medical practice is rife with cases in which, despite
empirical evidence that should have been obvious that things were going wrong, doctors stuck with their old practice,
rationalizing it on traditional grounds. Part of my analysis will be concerned with conditions that make it possible for
effective learning to proceed, for there are conditions under which this is virtually impossible. Part will be concerned
with the nature of effective learning processes under those favorable conditions, but also with obstacles to productive
learning.
This theoretical perspective on problem solving, which sees the actor as both theorizing and engaging in trial and
error learning, is similar in some ways to the models developed by Gavetti and Levinthal (2000) and by Gavetti (2005).
However, my approach here puts special emphasis on the role of trial and error learning in potentially leading to
improvements in the theory held by the actor.
In my analysis, I highlight a key variable often neglected in theories of problem solving—the ability of the problem
solver to see and recognize the attributes of paths that enable distinguishing among them in a relatively fine-grained
way, and to see and associate outcomes reliably with path attributes. These abilities are essential if one is to be able to
develop or revise a theory, a map to guide further exploration, based on what one has learned through action or through
empirical research. Limitations on these abilities may limit the possibilities of learning, and their improvement may be
the key to developing a useful theory of what works and what does not. Thus in many fields, medicine in particular, the
basis for a significant improvement in practice often is laid by enhanced ability to make fine-grained discriminations,
or to see things that previously were unseen, which enables the development of more accurate understanding (theory)
of what is going on, that in turn leads to better practice.
Recognition that both theorizing and trial and error learning are involved leads one to see concepts that are widely
invoked in the literature on problem solving in a somewhat different light. Thus within this framework the “local” of
“local search” needs to be understood as theoretically or cognitively local. Similarly, whether a topography is smooth
or rugged depends on our problem solver’s cognitive mapping. These phenomena are clearly recognizable in some
case studies of actual problem solving, but seem not to have been effectively flagged in the theoretical literature of
search under bounded rationality.
Another conceptually interesting, and very real, phenomenon that becomes clear under the formulation I propose
is that theories need not be causally valid to be useful in guiding action taking. On the other hand, the practical value
80 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89

of a theory that is not connected with the actual causal forces at work may be very context specific, and holding to
that theory may lead to disastrous choices if the context changes. Theories that have the causal factors basically right
are less likely to be vulnerable to such problems. Again the argument here clearly fits many empirical cases, but while
occasionally recognized, has not been highlighted in theoretical writings.
While I believe my formulation is a particularly apt characterization of the cumulative efforts to improve technology
in areas where, as in medical practice, over the years there has been significant progress, I also believe that it sheds
light on why it seems so difficult to improve practice cumulatively and significantly in a number of other arenas, for
example in education or, I would argue, business management. More generally, the formulation helps to illuminate
some of the key reasons why progress has been so much more rapid in certain fields of activity than others.
I proceed as follows. In Section 2 I lay out the general analytic formulation. In Section 3 I narrow the focus and
explore a highly stylized model, in order to develop concrete if abstract examples of some of the general analytic
arguments presented earlier.
Then, in Section 4 I broaden the analysis, adopt a less abstract style, and reflect on the factors that make problem
solving in a field hard, or relatively easy. The examples I use will be drawn mostly from the history of technology,
and most of these will be concerned with the advance of medical know-how. However, I also will draw contrasts with
search and problem solving in the fields of business practice and in education.

2. The general model

Consider the following general problem. Our actor faces a goal-oriented task that must be done periodically. Each
time he performs the task his objective is to get as close as possible to his goal. He faces a set of S paths of which he
is aware and believes are open to him, that vary in how close they get. How close a path gets to the goal is measured
by a figure of merit, M. The actor wants to pick a path with as high a value of M as is possible. His learning task is to
come to identify paths with high-expected M.
The task context may be a constant, at least for a considerable period of time, or it may be variable. If it is variable,
what paths are best may be a function of the environmental state. The actor’s learning task clearly is more complicated
under these latter conditions than when the task environment is a constant. He either needs to learn how to identify
different contexts, as well as a set of context-specific guides to action, or find a broad guide to action that works
reasonably well in all or most contexts he will face. To facilitate coherent development of the main argument, for
the moment I hold off consideration of discrimination among contexts, and assume that our actor proceeds unaware
of or unable to distinguish different contexts; later I will open up the analysis in this regard. Under these restrictive
assumptions our actors learning task, then, is to find a way of choosing paths of action, among those that he believes
open to him, that work well on average, given the possibly broad range of contexts in which he may have to operate.
Each possible path is marked by the values taken on by N characteristics or attributes. Some of these characteristics
may simply be present or not on a path; others may have a number of possible values. Our actor is able to see at least
some of these attributes, before committing to a particular path, but not necessarily all of them. At any time he has a set
of beliefs about what visible attribute values mark good paths (and may have some conjectures as to desirable values
of attributes he presently is unable to see, but in the future might learn how to).
When picking a path at any time our actor can only control that that path has the attributes he deems promising that
he is able to see. If there is more than one path that has the attribute values our actor is focusing on, which path actually
gets taken is determined randomly (equal probability) across the set of paths with those attribute values.
Thus what happens as a result of the actor’s choice is determined partly by his deliberate choice, and partly by
chance. It is as if a choice by the actor involved picking a particular urn, defined by the visible attributes chosen and
under control, and containing all paths with those markings, from which random draws are taken, with replacement (for
a more extensive development of the “urns” model, see Nelson, 1982). If our actor keeps choosing a course of action
with a particular set of controlled attributes, he will achieve a mean payoff. But there may be considerable variance.
Part of that variance may be due to the fact that the paths actually taken vary. Part may be due to the fact that there are
different context conditions that the actor does not notice.
In the analysis that follows I will assume that the actor’s choice of paths with a particular set of attribute values, a
particular urn, stems from a mental model or “theory” held by the actor. I do not want to be too constrained here in
the interpretation of what a “theory” means. A theory about good paths may have backing in scientific understanding,
or it may not. It may simply reflect the actor’s experience. The theory may be deep and provide an explanation for
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 81

why particular kinds of paths are better than others, or it may simply be about the nature of good paths. The empirical
evidence supporting the theory may be strong, or weak. The theory may be no more than a hunch. In any case, the
important thing about our actor’s theory is that it identifies the characteristics of paths that he thinks lead to good
outcomes.
As I indicated earlier, the analysis here is strongly motivated by my developing understanding of innovation in
medicine. From this orientation, the particular paths can be interpreted as different treatments for a disease, or at least
for a set of symptoms. Given those symptoms, our actor–physician’s choice problem is to come to find paths–treatments
that are likely to have a high figure of merit, given those symptoms.
As the formulation is set up, there clearly is a lot of room for learning. That learning may simply come as a byproduct
of doing, a physician learning over time what works and what does not work in the course of treating a series of patients
with symptoms that, at the start of the series at least, were considered to demark the same problem. However, there also
usually are opportunities to pick a path to follow in part at least because doing so will help test or fine tune prevailing
theory, or because such action will enable the exploration of a somewhat different theory. In the case in question, there
likely are significant opportunity costs in experimenting this way, involving major risks for the patient. The advantages
of being able to do such experimentation “off-line” are obvious, not involving actual treatment of patients until strong
evidence has been gained of the efficacy of the experimental treatment. Later in this essay I will distinguish between
search that proceeds off-line, as in a separate R & D activity, and on-line experimenting, and their often important
interactions. But in this and the following section I will repress these complications.
Search to improve prevailing theory or find a better one can be oriented in a variety of different ways. For example,
in the actor’s mind the objective of search may be to try to improve the theory by discriminating among different
path markers in a more fine-grained way. Or the relevance of markers not considered under prevailing theory may be
explored. Or, and here I loosen my earlier assumption, attention might be given to trying to distinguish among different
context conditions. Or, if our actor is entertaining more than one theory, search may be oriented to finding which one
is better.
Given a broad orientation to an effort to improve theory about what makes for good paths, there are a variety of
possible search tactics. I suggest that the concept of local or neighborhood search can be interpreted as a particular kind
of search tactic for fine tuning a theory. The concept is based on the notion of a starting point, under our assumptions
here associated with a particular path or set of them that have been taken as a result of a broad theory that has been
held by our actor, and then exploration as to whether some marginal change in orientation (which would amount to a
marginal change in the guiding theory) might increase expected M.
My use of the term “path” to denote a course of action suggests that geography matters. While clearly inappropriate
as a characterization of the relevant “topography” in, say a search for a pharmaceutical that has certain properties,
the idea of a physical geography is useful as a starting place for my conceptual discussion here. Thus a theory
about the characteristics of good paths might be that they tend to start in a particular geographical area. Under that
theory, geographical proximity is a natural way to think of closeness or neighborhood, and local search might involve
experimentally shifting search to an area that is physically proximate to the one that had been being employed as a
starting point.
But a theory about the characteristics of good paths may have nothing to do with the geographical location of the
start of the path. The broad theory may be that paths that go through dense woods tend to be good ones. And that
starting in a grove of trees is a good indicator that the path will spend a lot of its course in the woods. Under that
theory, our actor might think of attending the number of trees in the grove at the start of a path. Assume he gets close
to his goal after taking a path with say X trees in its starting grove, and recognizes that number. He might then be
interested in exploring whether paths with a few more or a few less trees in the grove at the start might yield even
higher expected M. Those paths might not be nearby geographically, but would be “nearby” in their markers. Local
or neighborhood search seems generally to be regarded as theory free. However, I am proposing here that notions of
nearness, and similarity, are very much influenced by the background theory the actor has.
Consider, for example, the concept of similarity of chemical elements. If a chemist is looking for an element that
has many of the same reactive characteristics as the one he presently is working with, but which may have fewer
drawbacks, he will be drawn to other elements on the same column of the periodic table, a similarity defined in terms
of the number of electrons in the outer shell. A physician prescribing for angina, who has a patient that has reacted
badly to a particular beta-blocker, likely will try another beta-blocker. In both of these cases, what is similar, what is a
local change, is highly “theory” dependent.
82 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89

Often the concept of neighborhood search is linked to the concept of hill climbing. The notion of hill climbing
assumes a space of alternatives that can be connected in a chain, so that it is feasible for the actor to move from some
initial point A, to some point in its neighborhood, say B, where the payoff may be higher, and if it turns out to be
higher, then move from point B to point C, which is in the same general direction from B as B was from A, and so on.
Like the concept of neighborhood search, hill climbing generally seems to be regarded as proceeding in a natural
physical space, and as a theory-free activity. But, following the line of argument above, I propose that, in problem
solving, hill climbing generally should not be conceptualized as proceeding in physical space. Thus our actor may
pose the question in terms of whether the number of trees at the start of a path matters, and move from two trees in the
grove paths to three tree paths to four tree paths, etc. It may turn out that, while merit changes erratically as one moves
from one path to another that is nearby physically, average merit is smoothly related to the number of trees that mark
the start of a path. What is a rugged topography and what is not, just as what is close and what is not, depends on the
theoretical lenses through which one views the terrain.
As the above discussion indicates, and as has been well recognized in various literatures, the channeling of vision
and of efforts at problem solving provided by a strongly held prevailing theory can be a plus, or a minus. Significant
progress may require that a new theory come into view, and the old one abandoned. Thus progress in dealing with peptic
ulcers was limited so long as the prevailing theory was that these were caused by acid in the stomach, and effective
remedies were those that reduced that acidity or shielded the stomach wall from the acid. When it was discovered that
the acidity was due to the workings of bacteria, a whole new theory of how to deal with the problem came into view,
one that was oriented towards the use of antibiotics (see Thagard, 1999). Similarly, a new way of preventing as well as
treating cervical cancers was brought into view by the recognition that the causal factor often was microbial.
The perspective I am developing here also flags attention to the fact that there may be several different theories
that are roughly consistent with the empirical evidence regarding the characteristics of good and bad paths. Of these
theories, some may work for quite spurious reasons, and be quite context specific.
Thus our actor’s theory that paths that go through the woods are on average better than paths that go over fields may
be supported by considerable experience. But some of the paths through the woods go by lakes and others do not. And
in fact it may be passage by lakes that gives a path its good properties. The reason that paths through the woods payoff
more on average than paths through meadows may simply be that a much larger percentage of the former than the
latter go by lakes. If this is learned, our actor might well be able to do better with this new and more accurate theory,
at least if there are visible indicators that are more reliable than a grove of trees at the start in telling whether a path
will go by a lake.
An interesting real example is the growing perception, during the 18th and early 19th centuries, among physicians
concerned with epidemics, that malodorous air, miasma, was an important source of contagious disease. That theory
led to campaigns to clean up the urban environment that, when actually carried out, helped to prevent epidemics. This
“theory”, which clearly contributed to good public policies, existed well before there was general acceptance of the
germ theory of disease (see e.g. Porter, 1997, Chapter 13). However, when the germ theory of infections disease came
to be accepted and understood, and discriminations made among different diseases and their carriers, policies could
become much more focused on preventing the presence of the carriers of various kinds of diseases.
As I suggested earlier, the achievement of a high average M by our actor may require the ability to recognize different
contexts and pick paths that are appropriate to each. A “by a lake” theory of the key characteristic of good paths may
work well in a context where there are lots of lakes, and no rivers, but not be helpful in a context where there are no
lakes. Conversely, if rivers have the same favorable properties as lakes, a “good paths go by a river” theory may be
fine in a context where there are no lakes and several rivers. If our actor always is in one context or the other, he only
needs to have one of these theories, the right one for the context. But if he sometimes is in one context and sometimes
in the other, he needs to be able to recognize the difference and respond to it correctly.
Of course, he might do fine if he had a general theory that what counts about paths is whether or not they go by
a body of water, and paths had visible markers that reliably signaled that. Coming to understand the general theory
sounds both more parsimonious and deeper. However, if there is no reliable general marker, knowing that good paths
go by some kind of body of water may tell our actor little about how to identify paths that go by lakes if he is in lake
country, or by rivers if he is in river country. At the same time, there might well be reliable markers for whether a path
goes by a lake in lake country, and others for whether a path goes by a river in river country, and ways that our actor
can tell where he is. In this case, holding the correct general theory may add nothing to our actor’s ability to choose
good paths.
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 83

I would like to argue that in many arenas of problem solving, even many that are illuminated by a formal science,
general theory often points only very broadly towards good practice. The devil is in the details, and understanding of
the details is won only through detailed empirical exploration guided only broadly by that general theory.
The germ theory of (many) diseases clearly was an important breakthrough, and greatly aided efforts to find ways
to deal with human disease. But there are many different varieties of infectious disease, and public health measures
that are effective for one may not do much to stop another. Cholera bacteria are carried in contaminated water, and
effective campaigns to stop cholera (and typhoid) focus on keeping the water supply clean. Malaria is carried by
mosquitoes, and effective public health programs are focused on this vector. A general germ theory motivated search to
identify the particular disease vectors involved in particular diseases. But the usefulness of the theory in guiding public
health measures depended on the success of fine-grained research. As noted above, in some early cases effective public
health measures were stumbled on without the benefit of the germ theory. However, the emergence of a germ theory,
combined with the development of ways of “seeing” and identifying the guilty microbes, greatly enhanced ability to
make progress on this class of problem.

3. A specific abstract example

I present now a highly abstract version of the general analysis sketched above, for the purpose of making clearer
some of its implications. We observe this particular example from a position of full knowledge of the problem facing
our actor, so we can appreciate his difficulties with the problem, see what can lure him to false conjectures, and the
ways he might be able to develop a good theory and course of action. In this particular case it is assumed that the
context facing our actor is a constant, or at least is a constant over the period of time we consider in the formal analysis.
On the other hand, one of the questions we explore is the relevance of what the actor has learned in this context if the
context changes.
As displayed in Fig. 1, there are 36 actual different paths of action. No new ones are introduced or discovered
over the course of the analysis; there is no technological innovation possible in that sense. The merit of each available

Fig. 1.
84 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89

pathway is shown in the upper left hand corner, which we can see but the actor can’t until he has gone down that
particular path. In this particular model I assume that once a path is traversed, the actor has no trouble assigning a
merit rating. His problem is in knowing what particular path he actually took. He only knows certain characteristics or
signs of that path. Thus what he can see influences his ability to control what he actually does. For this reason, among
others, just what he is able to see matters a lot.
I, the author, have laid out the alternative paths on an orderly grid which you, the reader, can see. However, our actor
has no map and has no compass, and location in Euclidian space is something he cannot fathom. On the other hand,
he can see, if he pays attention, that paths, or their starting points, differ regarding whether they have an X mark in the
middle, or a rail line like mark running across the bottom, or neither of these. There are other features, or finer features,
that differentiate the alternatives, and these are displayed in Figs. 2 and 3. But at the start of the problem-solving
venture in question, our actor cannot make out these differences.
The reader may verify that the average merit from picking a path at random is 1. If our actor started by randomly
picking paths, it might take him some time before he pinned this down, because there is considerable variance. However,
perhaps the variation in achieved merit in itself would start him thinking about whether certain identifiable features
might matter, for example an X mark or a rail line mark. If he explored one or another of these possibilities, he would
find out that those variables did matter. Paths marked by an X have an average merit of 15/6; paths with rail lines have
on average a merit average of 21/8. Either of these is a useful theory to guide choice, in this particular context at least.
And in this context, our actor might find it difficult to decide which of these theories is a better guide to action, because
there is not much difference between the average merits of the paths pointed to by the two theories.
This is interesting because, as will become evident shortly, the “marked with an X” theory of high merit paths
is completely spurious causally, and works in this context only by chance. It likely would be of no use at all if the
context changed. In contrast, the “marked with a rail line” theory does have a relationship with the underlying causal
mechanisms.

Fig. 2.
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 85

Fig. 3.

I note that, in the spirit of the discussion of the prior section, if our actor can only tell whether a path has a rail line
like marker, or an X, he has limited control over the path he actually takes at any time. A consequence is significant
variation in the values of M achieved from taking paths that, for all our actor knows, are the same.
Fig. 2 displays what our actor can see after he gets eyeglasses that enable him to see more clearly. One thing he
now can see is that, while he missed it without glasses, some of the paths marked by an X also have an x, while others
do not. Being able to make this discrimination clearly is helpful, at least in this context. Paths marked by an X but
without an x have an average payoff of 11/3. If he can find this out by experimenting, and sticks to paths so marked,
this beats the average he can get if he guides his choices by whether a path has a rail line mark or not. However, while
a theory that good paths are marked by a simple X will help guide him in this particular context, the theory is spurious
regarding true causality.
Good glasses also enable our actor to see more finely into the markers on paths that he used to think were characterized
by rail line markers. He now can see that these really are a set of connected crosses. He also can see that the set so
marked seems to differ in terms of the number of crosses or cross bars. What he can make of that knowledge depends
on how finely he can discriminate. If he can actually count the number of crosses or cross bars, and experimented to
find out the relationship between average merit and number of crosses, he might be able to home in on paths with seven
crosses, which have an average merit of 13/4, not quite as high as the paths marked by a simple X, but pretty close.
However, it would not be easy to reach that understanding. There are a large number of paths marked by collections
of crosses, and several for each number. Attempts at hill climbing through finding the average merit for paths with a
given number of crosses and then moving up or down one would not work very well. First of all, there still is more than
one path associated with any marker our actor can identify, and therefore different values of M in going down paths
86 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89

of that class. At this level of discrimination, it is average M that counts, but getting the averages close to right might
take a number of trials. Second, the relationship between average merit and number of crosses is far from smooth. The
topography is rugged, when looked at from the perspective lent by a theory that focused on the number of crosses.
Now equip our actor with an even better set of glasses. He now can see, as shown in Fig. 3, that some of what
he thought were crosses on different paths in fact are asterisks. He could then explore the merits of the paths that
had some asterisks. If he had any luck, he would quickly come to a theory that asterisks marked good paths. The
average merit of paths with asterisks is 21/6. If he reached this conclusion, he might then consider whether the number
of asterisks mattered. This would be an easier task than exploring whether the number of undifferentiated cross-like
things mattered. There is only one path that has a particular number of asterisks; no intra class variation to make things
complicated. That is, finer discrimination gives the actor better control of what he actually does. And here hill climbing
works. The topography is smooth. Merit goes up linearly with the number of asterisks.
In fact, this is the only structural systematic relationship between attributes and merit in the set of paths. I know
that because I built the model. There are six paths with asterisks. In each of these merit equals the number of asterisks.
The merit for the other 30 paths is either zero or one, half each. Any relationships between attributes other than the
number of asterisks and merit are due to chance, or to the fact that certain attributes are partially tied to the number of
asterisks. The particular context characterized in Fig. 3 was achieved by scattering the paths so characterized within
the grid more or less at random.
It is useful to reflect a bit on the robustness of the different theories that, in our account, the actor used to guide his
choice of paths, all theories that, in this particular context, helped him to achieve an average level of merit that was
better than chance. The theories relating to the presence of an X, which our actor could see without glasses, or the
theory about the high average merit of paths marked by an X but without an x, which he could see with glasses, are as
I stated earlier completely spurious causally. They are helpful in this particular environment. But lay out the random
elements of the grid in another way, and guidance by such theories is likely to be worthless, or worse.
The “marked by a rail line”, or connected cross-like things, theory will provide some guidance in another environ-
ment, because all paths that are marked by asterisks also would be seen as being marked by a rail line or connected
crosses, if the actor had no glasses or weak ones. And in most contexts, having a large number of cross-like things
likely would be a guide to good paths, because paths with a large number of asterisks are in fact good paths.
With the vision of hindsight, you (the reader) know that the true guiding theory is simple. Merit equals the number
of asterisks for those paths so marked, otherwise zero or one with equal probability. If our actor arrives at that theory,
he is in good shape even if the context changed. However, note that within the context we have just considered, it took
the development of ability for fine discrimination for our actor to have any possibility at all to arrive at the correct
general theory; without that capability while our actor might develop a theory that would enable him to do well in the
context in which he is operating, there is no way that he can achieve the correct theory of what determines merit. And
even if he had high-quality glasses, it might be very difficult for him to find the correct general theory. Among other
things, in this context there are other theories that provide almost as good guidance.
But if the context changed, how well our actor would do would depend very much on the theory he had developed
in the earlier context. If he had come to the correct general theory, the change in context would not affect his ability to
do well, even in the short run. Of course, even if he did not hold the right general theory, with time he might be able
to come to a theory that effectively guided action in the new context. But if the context changed again. . .

4. On factors that enable progress in problem solving

In this section I lower the level of abstraction and reflect on what enables difficult problems ultimately to be solved
reasonably well. My focus here is on persistent problems relevant to a significant group of individuals or organizations,
on which progress is made not all at once, but rather from a series of partial solutions that enable the goal to be
approached more and more closely. This aptly characterizes the advance of most technologies.
I want to start by proposing that the ability to see fine-grained differences, and to control the path one takes, what
actually is going on, based on fine-grained discriminations, a key variable influencing the power of search highlighted
in Section 3, shows up as important in many empirical studies. While the point made there would seem obvious,
almost trivial, there is much more than tautology here. Ability to see, discriminate, and reliably select and control that
the actions have particular characteristics, is something that cannot be taken for granted. In many cases creating the
capability to discriminate and control has been the key to becoming able to solve the problem.
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 87

Consider, as a famous example, the inability of navigators of ships at sea to assess accurately their longitude, prior
to the development of the chronometer. As a result, those responsible for choosing the ship’s path often could not judge
accurately or control the longitudinal dimension of the path they were taking, with occasional disastrous consequences.
The development of the chronometer enabled navigators to have much finer control over their ships’ actual path, given
the course they aimed to take (for the story see Sobel, 1996). In effect the “subset of paths” on which they might be at
any time was substantially reduced.
This was a case where, while improved ability to see and control greatly improved ability to pick and hold to
good paths, these developments had no significant affect on broad theory. The importance of being able to mea-
sure longitude had been understood for a long time; only the ability was lacking. However, in many cases such
improved abilities have led to new theories, or at least permitted them, which in turn led to the discovery of better
practice.
Thus prior to the development and acceptance of the microbial theory of infectious diseases, largely as a result
of the research and publicity done by Pasteur and Koch, physicians had little clue regarding how to deal with such
illnesses. Efforts to cure individual patients usually were worthless or worse. With the acceptance of microbe theory,
this changed. But microbial theory could not have been discovered or demonstrated (although it had been conjectured)
without instruments and techniques that enabled microbes actually to be seen. In turn that theory, and the observational
capabilities that enabled it, set problem solving in the direction of identifying the particular microbes involved in
various diseases, again a path requiring fine-grained perception and control. Once the particular micro-organism was
identified, problem solving could turn towards trying to develop an effective vaccine. While a persuasive theory of
why vaccines often worked only came later, experience gave some strong clues as to how to develop a vaccine, once
the offending microbe had been identified. And while risky, tests of the efficacy of a new vaccine generally gave sharp
results. The germ theory of disease also provided useful focus for efforts to find or create antibiotics, but significant
progress in this area took a long time to achieve.
There was an interesting difference between the ability to treat individual patients and ability to devise public health
measures for infectious diseases, prior to germ theory. The dominant theory that guided treatment of individual patients
by many physicians was based on the notion that illness was the result of the balance of body fluids getting away from
what it should be. While for some physicians this theory led to advice to patients regarding a healthy life style, it also
often led to treatments for illness like bleeding, that usually did more harm than good. In contrast, as I noted earlier
there was a relatively widely held theory about the health problems caused by “bad airs” that led to various campaigns
to clean up areas that were malodorous. In some cases this led to the elimination of various public health hazards, and
generally led to actions that did more good than harm. This is a good example of an invalid theory that in fact enabled
some effective problem solving, because the factors it identified as relevant were often correlated with the real causes,
and dealing with the correlated variables often dealt with the actual culprits. Of course with germ theory, public health
policy would rest on a new and more effective foundation.
I have been arguing that one of the important advantages of being able to identify, specify, and control closely the
paths of action that are taken is that one then has the possibility of getting useful feedback regarding efficacy. But while
ability to fine tune action may be necessary for sharp feedback, it is not sufficient. In the model in Section 3 our actor
had no difficulty assessing merit when he got to the end of a path. However, there are a wide variety of problem areas
where, while actions can be tightly specified and controlled, feedback is not sharp or rapid. This is often the case in
medicine.
Areas of medical practice differ significantly in this respect. In general, one can learn relatively quickly if a new
antibiotic works or not, although it may take more time to learn about possible side effects. However, with treatments
that are basically intended to prevent premature deaths, which is usually the case with treatments for early breast or
prostate cancers, it can take a long time before one has statistically significant data on the ages of death of a treated group,
and a control group. This very slow feedback regarding efficacy clearly adds to the difficulty of making significant
progress in these areas.
Another striking feature of progress in medicine, and the evolution of technologies more generally, that is repressed
in the model, is that achieving a goal, or a continuing improvement in performance, almost always is the result of work
done by many different parties. There are exceptions. Jenner’s bet on the efficacy of using materials from cowpox as
the stuff for vaccination for small pox was based on his access to folk wisdom, but the idea of the vaccine was his.
Similarly, Pasteur’s vaccine for rabies was largely his doing. But in many cases the development of a successful means
to prevent or cure a disease has been the result of the work of many different people and organizations.
88 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89

The penicillin case is a good example. Here the original discovery of the antibiotic properties of the mold was
done by one person, Fleming. The experimental work on using the substance as a treatment for infection was done by
someone else, Florey and colleagues. And the development of an effective production method was done by still other
people. There are many other similar examples in the history of medicine. In virtually all detailed studies of particular
important inventions I know about, a history of earlier work that laid the basis for the invention in question is part of
the story.
Again, I want to propose that a necessary, if not sufficient, condition for the ability of today’s problem solvers to
build on the work of yesterday’s is the ability to closely specify and control the paths being worked on. You can not
build on someone else’s work unless you know the nature of that work in a certain amount of detail, and have the ability
to pick up where that prior work ended.
In Sections 2 and 3 I ducked the question of just how exploration for better ways of doing things proceeded, while
noting that it could proceed on-line or, in some cases, off-line. Ability to experiment, to learn, off-line obviously has
major advantages. Doing experimentation on-line involves opportunity costs, if a path other than in the set currently
judged most promising is taken in order to explore its merit. If one tries to keep opportunity cost down by not diverging
much from known best practice, the range of exploration is seriously constrained, and what one learns has to come
largely as result of “learning by doing” where ability to pin down cause and effect relationships may be quite limited.
In most of the modern technologies where advance has been rapid and sustained, much of the exploration and fine
tuning of new paths is done off-line, in R & D that is separated from actual practice.
The fact that today new medical practices, including the use of new pharmaceuticals, and devices, are largely
developed off-line is so familiar that it is not recognized as remarkable. Until recently, all advances in medical practice
were achieved in exploration on-line. Jenner’s discovery of vaccination is a good example. While the discovery that
the availability of citrus fruits could prevent scurvy involved a relatively controlled experiment in which different
sailors were provided with different supplements to their standard diet, that experiment was very much on-line. As
this example illustrates, it always was possible to experiment on a small number of people, with broader practice for
the population as a whole learning from advertised experimental results. However, this is very different from doing
experimentation off-line, as it were, with application to even a small number of patients dependent on “laboratory”
results.
Indeed, prior to the mid-19th century there was hardly any “laboratory” work that led to significant advances
in medical practice. A combination of the development of the germ theory of disease, identification of cells as the
basic building blocks of living creatures, and significant advances in understanding of biochemistry, changed all that.
Nowadays it is expected, indeed required by law, that a new treatment, a pharmaceutical for example, be sufficiently
well tested off-line, in the laboratory, or using nonhuman subjects, before actually trying it on humans, that the prospects
that the new treatment will be effective are relatively good and the chances of its causing harm relatively small.
Let me flag what by now should be obvious to the reader. The efficacy of off-line R & D depends on the ability to
specify closely, to control, what is achieved in R & D. And it must be possible to take what has been learned off-line,
and use it on-line with much the same results as was achieved in R & D. As Nightingale (2004) has argued, much of
what is achieved in R & D is achieved under tightly controlled conditions. Thus the transfer to practice often involves
developing shields so that the relevant aspects of practice resemble the controlled conditions of the R & D setting.
Reflect on how much shielding from possible adverse environmental effects goes into the design of high tech products,
like a micro-processor. Pharmaceuticals are protected by an outer shell, and conditions of use closely specified.
I want to conclude this section by pointing to two areas of practice where progress has been very slow, and a central
reason (in my view) is that the operating practices are very difficult to pin down in any detail (for a more extensive
discussion see Nelson, 2003). One is education. Consider as an example the “phonics” method of teaching reading. The
broad outlines of the phonics method are relatively clear. However, few schools or teachers use it in a pure form, and
more generally the details of how reading is taught under its broad rubric differ from school to school, teacher to teacher,
and even student to student. For this reason, among others, it is very hard to evaluate the efficacy of using phonics,
as contrasted with other philosophies of teaching reading that march under a different flag. This problem regarding
educational techniques has been well recognized, and it has been proposed to establish tight standards for particular
techniques. But this is easier said than done, partly because it is recognized that the imposition of tight standards might
in many cases eliminate the possibility of context-specific variation that is needed for efficacy in particular cases. But
to the extent that broad flexibility is needed, what is learned though experiment and experience is going to be coarse
grained.
 R.R. Nelson / Journal of Economic Behavior & Organization 67 (2008) 78–89 89

I would like to propose that the same thing is true of much of business practice. The M form, quality circles, just
in time, all are terms that cover a lot of variety. Part of that variety is intentional, involving tailoring to fit particular
contexts, but a good part stems from the fact that the individual actors involved are not clear themselves regarding
exactly what they are doing. This makes for a situation in which learning, individual or collective, is not easy.
It is noteworthy that off-line empirically oriented R & D do not provide much useful knowledge about what works
and what does not in either of these areas. That is partly because just what is being tested in R & D usually is not easy
to specify exactly. It also is because, even if it could be tightly specified, and brought to on-line practice exactly is it
was in R & D, performance in a tightly controlled setting may tell very little about what performance would be on-line.
So in these areas of human activity learning by doing is difficult and learning through R & D harder still. I propose
that the broad theoretical structure developed in this paper tells why.

Acknowledgements

The author is grateful to Giovanni Gavetti, Annetine Gelijns, Paul Nightingale, and Sidney Winter for helpful
comments on early versions of this paper, and to the Sloan Foundation, and the Merck Foundation, for research
support.

References

Clark, A., 1997. Being There: Putting Brain, Body, and World Together Again. MIT Press, Cambridge, Massachusetts.
Dosi, G., 1982. Technological paradigms and technological trajectories. Research Policy 11, 147–162.
Dreyfuss, H., Dreyfuss, S., 1986. Mind Over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. Free Press, New
York.
Gavetti, G., 2005. Cognition and hierarchy: rethinking the microfoundations of capabilities development. Organization Science 16, 599–617.
Gavetti, G., Levinthal, D., 2000. Looking forward and looking backward: cognitive and experiential search. Administrative Science Quarterly 45,
113–137.
Hendriks-Jansen, H., 1966. Catching Ourselves in the Act: Situated Activity, Interaction, Emergence, Evolution, and Human Thought. MIT Press,
Cambridge, Massachusetts.
Holyoke, K., Spellman, B., 1993. Thinking. Annual Review of Psychology 44, 265–315.
Johnson-Laird, P.N., 1983. Mental Models, Towards a Cognitive Science of Language, Inference, and Consciousness. Harvard University Press,
Cambridge.
Leonard-Barton, D., 1992. Core capabilities and core rigidities: a paradox in managing new product development. Strategic Management 13,
111–125.
Nelson, R., 1982. The role of knowledge in R and D efficiency. Quarterly Journal of Economics 96, 453–470.
Nelson, R., 2003. On the uneven evolution of human know-how. Research Policy 32, 909–922.
Nelson, K., Nelson, R., 2002. On the nature and evolution of human know-how. Research Policy 31, 719–733.
Nelson, R., Winter, S., 1982. An Evolutionary Theory of Economic Change. Harvard University Press, Cambridge.
Newell, A., Simon, H., 1972. Human Problem Solving. Prentice Hall, Englewood Cliffs.
Nightingale, P., 2004. Technological capabilities, invisible infrastructure, and the un-social construction of predictability: the overlooked fixed costs
of useful research. Research Policy 33, 1259–1284.
Porter, R., 1997. The Greatest Benefit to Mankind. Norton, New York.
Sobel, D., 1996. Longitude: The True Story of the Lone Genius Who Solved the Greatest Scientific Problem of His Time. Penguin, London.
Thagard, P., 1999. How Scientists Explain Disease. Princeton University Press, Princeton.
Tripsis, M., Gavetti, G., 2000. Capabilities, cognition, and inertia: evidence from digital imaging. Strategic Management 21, 1147–1161.
Vincenti, W., 1990. What Engineers Know and How They Know It. Johns Hopkins University Press, Baltimore.