The Ambiguities of Experience, page 3
Two Sets of Questions
The literature on processes of adaptation that replicate success addresses two related, but different, sets of questions.
The first set includes: What do individuals or organizations do in this situation? How do they act? How do their actions change over time as a result of experience? To what extent, in what way, and at what rate do they respond to knowledge of experienced results? To what extent do they learn to pursue the best alternative?
The second set of questions includes: What should an intelligent person or organization do in this situation? How long should several alternatives be sampled in order to obtain information about them? When should a choice be made and on what basis?
With respect to the second set of questions, mechanisms for the replication of success share a fundamental problem that is also conspicuous in calculative rationality—indeed, in all adaptive processes: How do they recognize and achieve an optimal balance between exploitation and exploration? (See Kuhn 1962, 1977; Holland 1975; March 1999c, chap. 7; Chen and Katila 2008; Fang and Levinthal 2009.) Exploitation refers to the utilization and refinement of what is known. It is reflected in efforts toward efficiency, standardization, accountability, and control. Exploration is the pursuit of what is not known. It is reflected in efforts to generate and experiment with deviant procedures and new possibilities. In processes involving the replication of success, the problem is typically a problem of allocating resources between efforts to learn more about the world (exploration) and efforts to take advantage of what is already known (exploitation). A frequent question about the replication of success is whether, as it responds to immediate feedback, it allocates too few resources to exploration (Starbuck, Greve, and Hedberg 1978; Miller 1994).
There are some obvious criteria that can be applied to evaluate the effect of replicating success on any particular set of rules for action:
Improvement: Does average performance improve with experience?
Stability: Does the likelihood that a choice at time t repeats the choice at t-1 increase with experience?
Reputation error: Is the past realized performance (reputation) of the chosen alternative greater or less than its expected value? How does the error change with time?
Optimality: Is the best alternative discovered and adopted? How long does it take? Or alternatively, how close is the chosen alternative to the optimum in expected performance over time?
COMPLICATIONS IN SUCCESS
REPLICATION
Trial-and-error learning, imitation, and selection have different properties as adaptive mechanisms, but they have a set of basic structural elements in common. These shared elements and the complications they create are the primary foci of the present discussion. For the most part, the phenomena considered here stem from properties of the learning environment and the adaptive mechanisms, not from any distinctive features of the cognitive apparatus that may be brought to bear on them. Thus they are different from, and substantially independent of, well-known cognitive limitations of individuals. These “structural” complications can be understood, but an understanding of the complications by an adaptive agent does not eliminate the difficulties.
First, history is complex. Even though the world may be orderly in the sense of following immutable laws, it is filled with complex causal relations. Forming correct inferences about history in the face of such complexity requires a complicated experimental design, a multivariate model, and a large sample. Unfortunately, the conditions underlying the replication of success in real experience include simple implicit experimental designs, simple implicit correlational models, and small samples. As a result, the replication of success is subject to extensive elements of misspecification and superstition.
Second, history is subject to stochastic uncertainty. The orderliness of the world is obscured by probabilistic variations. Identifying the best alternative through experience involves untangling the joint consequences of signal, noise, and sample size. Signal: The greater the true differences among alternatives, the greater the chance of identifying the true optimum by observing a sample. Noise: The smaller the stochastic variation in the observed outcomes, the greater the chance of identifying the true optimum by observing a sample. Sample size: The larger the sample, the greater the chance of identifying the true optimum by observing a sample.
Since experience in organizations often suffers from weak signals, substantial noise, and small samples, it is quite likely that realized history will deviate considerably from the underlying reality. Adaptation responds, not to the distribution of potential histories but to the specific history realized in a small sample. The results may lead to a more favorable experience with an alternative than is warranted, thus leading to a mistaken replication. Alternatively, the results may lead to a less favorable experience than is warranted, thus leading to a mistaken avoidance of replication.
As is well known to students of stochastic processes, stochastic variation produces some quite striking, counterintuitive surprises (Feller 1968). Many of these surprises are variations on so-called first-passage theorems or competitive maximum theorems. An example of the former is the case of coin flips: If a single fair coin is flipped a large number of times, what is the distribution of length of runs in which more than half of the results are heads? The usual intuition is that a fair coin should lead to relatively short runs of dominance by either heads or tails; the actual result is that the runs are, on average, rather long.
An example of competitive maximum theorems is provided by comparing two identical competitors, each of whom realizes each period a draw from a normal distribution with mean=0 and variance=1. Suppose we consider the average realized performance of the two competitors over time. As experience accumulates, the likelihood increases that the competitor with the greater average realized return at time t will also have the greater average realized return at t+1. That likelihood becomes quite high as t increases. The rankings of two competitors, in terms of average performance to date, persist for an extended time even though the two are identical in capabilities and even though the difference in average performance declines.
Third, the outcome possibilities for the various alternatives are affected by the sequence of choices made and the outcomes realized. Replication of success naturally affects the alternative chosen and thus the distribution of possibilities from which an outcome is drawn. Less obviously, it also often affects the distribution of possibilities for a given alternative. In standard terminology, the individual outcome distributions are endogenous to the choice or the realized outcome.
Some of the endogeneities can be characterized as cases of depletion: that is, the replication of success results, on average, in a lower performance because the replication itself is detrimental to subsequent outcomes. The most obvious cases involve resources that are depleted by use or competition. Other cases involve the corrosion of advantage through exploiting it (Barnett and Hansen 1996). Suppose, for example, that a tennis player follows a strategy of hitting the ball to the weaker side of his or her opponents, thus increasing the short-run prospects of victory but at the same time providing more practice for opponents in using their weaker hands than in using the stronger. Over time, practice effects will reduce the differences in competence between the two hands of competitors and thus reduce the competitive advantage of hitting to the weaker side. Other examples include cases of boredom or cynicism. Still others involve the adjustments of others, as for example when crying wolf changes the longer-run likelihood of a response to the cry.
Of possibly even greater importance, however, are cases involving outcome distributions that are augmented by use—where the replication itself improves the outcome distribution. An important case involves the effect of practice on performance. Each time a given alternative is chosen, its capability improves. It seems reasonable to assume that, typically, the effect of practice is to increase the mean and decrease the variance of the outcome distribution for the chosen alternative.
Practice effects complicate the use of success replication in finding an optimum choice. This complication is usually described in terms of “competency traps” (Levitt and March 1988; Arthur 1989). Suppose performance in a particular activity is a product of a fixed potential for the activity and a variable level of competence at it. Competence is characteristically low initially but increases through practice. Since performance reflects the product of competence and potential, practice effects make identification of the alternative with the highest potential more difficult than it would be in the absence of practice effects. It is quite possible that an alternative of lower potential will come to dominate one of greater potential by virtue of greater current competence on the former. This is particularly likely when comparing an existing inferior alternative at which an organization has a prolonged history of practice with a new superior alternative with which the organization is relatively inexperienced. Replication of success is more likely to aggravate this problem than to ameliorate it.
Similarly, if success makes subsequent success more likely, as in the “Matthew Effect” (Merton 1968), outcome distributions are endogenous to choices. Suppose, for example, that realized performance at time t, rt, is a draw from a normal distribution with mean=xt and standard deviation=st. If xt is a function of rt-1 (for example, xt=rt-1) the process takes on martingale characteristics, with the result that small initial differences among alternatives are converted, over time, into large differences. An obvious organizational example is the way in which evaluations of early performance influence evaluations of subsequent performance so that small initial variability in evaluations of personnel becomes large variability. A related example is found in the way that the likelihood of imitation of a particular practice depends on its “legitimacy,” which in turn depends on the number of others who have already adopted the practice (Carroll and Hannan 1989; Hannan 1998).
The replication of success is also affected by the definition of success in terms of a relation to aspirations and the ways in which aspirations are affected by achievements (Payne, Laughhann, and Crum 1980, 1981). Suppose the aspiration level at time t is a mix between aspiration at t-1 and performance at t-1. Thus, aspiration tends to track performance (is, in fact, an exponentially weighted moving average of performance) and essentially discounts current performance by a positive function of past performance. This makes subjective success or subjective failure particularly sensitive to noise in outcome determination. By making success (and thus the replications of success) depend on the history of past performances, adaptive aspirations tend to slow learning, especially if the rate of aspiration adjustment is rapid (March and Shapira 1992). On the other hand, if aspirations do not adjust to experience, so that success or failure tends to be stable, learning can become superstitious (Lave and March 1975).
Thus there are three simultaneous elements of learning. The first is learning what to do: looking for a good (or best) alternative technology, strategy, partner, etc. The second is learning how to do it: refining and improving competence with an alternative. The third is learning what to hope for: modifying the aspiration level for performance. The simultaneous adaptation of these three elements complicates the effectiveness of locating the best alternative. Adaptiveness in making a technology work and adaptiveness in aspirations interfere with adaptiveness in choosing a superior technology.
Fourth, sampling rates of experience are affected by sample outcomes. Experience in a particular time period can be seen as a draw from the distribution of possible outcomes associated with an alternative. Any particular experience is likely to be misleading to the extent to which there is variation in possible outcomes; and small samples of experience will have greater sampling error than will larger samples. The repetition of alternatives associated with success and the avoidance of alternatives associated with failure assure that the sample size of experience with successful alternatives will be larger than the sample size of experience with unsuccessful alternatives. As a result, the sampling error associated with unsuccessful alternatives will be larger than the sampling error associated with successful alternatives.
Two types of errors in experiential learning stem from sampling errors. The first is the error produced when a sample of experience has unrepresentative high returns. The second is the error produced when a sample of experience has unrepresentative low returns. Since the sample size of successful alternatives is increased by the repetition of success, the errors made by overestimating the value of an alternative tend to be self-correcting. Repetition reduces the sampling error associated with experience with a successful alternative, thus exposing errors of overestimation. On the other hand, errors made by underestimating the value of an alternative are not self-correcting. Alternatives that are better than their early results will tend to be persistently underrated and underchosen.
Some of the phenomena can be illustrated by a simple model, as long as it is recognized to be a very stylized representation. Suppose a choice is made each time period among a set of fifty alternatives. Each alternative (Ai) is characterized by a normal outcome distribution with a mean=xi and a standard deviation=s. The xi’s and s are fixed over time. The xi’s are themselves draws from a normal distribution with a mean=0 and a standard distribution=S. Each alternative has an initial reputation, Ri,0=0, the mean expectation within the population of alternatives. Subsequently, the value of Ri,t for each chosen alternative equals the mean realization associated with a choice of that alternative. Thus, reputations at t may be based on as few as zero observations or as many as t-1. Each time period the alternative with the highest reputation, max Ri,t, is chosen and an outcome realized (a draw from the distribution for that alternative).
The properties of this simple model illustrate the consequences of the endogenous sampling rate. Improvement: When the alternatives have different means (S>0), the replication of success improves performance over time. Better alternatives are discovered and replicated. Stability: There is a strong tendency for the replication of success to become very stable in its choice of alternatives. This result occurs, though with less force, even when there are no differences among the alternatives (S=0). Reputation error: The difference between the reputation of the chosen alternative and the true mean of the outcome distribution for that alternative (Ri,t-xi) is positive—that is, reputations of chosen alternatives overestimate capabilities. The error is large initially but declines over time, ultimately (after a very large number of periods) approaching zero. Optimality: If we compute the ratio between the mean of the chosen alternative and the mean of the best possible alternative among the fifty alternatives where S>0, the mean fraction of the optimum that the process realizes increases over time, but it falls far short of 1.0. The process rarely discovers the optimum.
A mixed story. The result is a mixed story. In simple situations where the causal structure is not complicated, replication of success frequently leads to improvement in performance over time if there are differences among the alternatives, their outcomes are relatively reliable (low variance), and an adequate sample of experience is obtained. Replication of success usually leads to stability, a steady increase in the likelihood of repeating a choice. It usually leads to improved reputation (past performance) of chosen alternatives for an extended period. Part of the improvement in reputation is due to the likelihood of choosing a better alternative, but part of it is due to selection of positive sampling error.
On the other hand, learning through the replication of success has troubling unfavorable properties. Even in simple situations, the choices made through replication of success are very likely to be substantially suboptimal. The replication of success at one level of learning confounds the replication of success at another level, producing, for example, competency traps. Even though the deviation of average performance from the expected value of the alternative ultimately approaches zero, for an extended period the realized performance of the chosen alternative reflects a substantial overestimation of that alternative’s potential. Even when there are no differences among the alternatives (and thus nothing to learn in terms of having a preference among them), replication of success usually leads to increased stability of choice. The subjective sense of learning is likely to be profound even when there is nothing to be learned.
A striking feature of these results is the extent to which they depend less on attributes of the learner than on attributes of experience. When experience unfolds in a way that makes learning effective, intelligence is augmented by the replication of success. But when experience is organized, as it often is, by complexity, ambiguity, stochastic variability, and limited sample sizes, the replication of success—whether through trial-and-error learning, imitation, or selection—is likely to lead to suboptimal outcomes.
LOW-INTELLECT LEARNING AND
HIGH-INTELLECT EXPLANATIONS
Although it is flawed in important ways, replication of success is a ubiquitous instrument of learning. In one form or another, and despite its substantial disabilities, it characterizes much of the adaptiveness of human actors. At the same time, however, the low-intellect simplicities of trial-and-error learning, imitation, and selection conflict with high-intellect hopes. Human conceit (both among actors and among observers) often seems to eschew attributions of human behavior to success replication in favor of more complicated, cognitive comprehensions, explanations, and justifications.
The joint ubiquity of success replication and of preferences for high-intellect explications of history and learning suggests the possibility that behavior that is commonly described in high-intellect terms actually may reflect rules learned through low-intellect replications of actions associated with success. It is an old idea, much beloved by Pavlovians and behavioral psychologists and (in a different form) by economic theorists.
