The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 99
The Economics of Artificial Intelligence Page 99

by Ajay Agrawal


  tal data on bargaining and on risky choice. The second idea is that some

  common limits on human prediction might be understood as the kinds of

  errors made by poor implementations of machine learning. The third idea

  is that it is important to study how AI technology used in fi rms and other

  institutions can both overcome and exploit human limits. The fullest under-

  standing of this tech- human interaction will require new knowledge from

  behavioral economics about attention, the nature of assembled preferences,

  and perceived fairness.

  24.2 Machine Learning to Find Behavioral Variables

  Behavioral economics can be defi ned as the study of natural limits on

  computation, willpower, and self- interest, and the implications of those

  Colin F. Camerer is the Robert Kirby Professor of Behavioral Finance and Economics at

  the California Institute of Technology.

  For acknowledgments, sources of research support, and disclosure of the author’s material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14013.ack.

  587

  588 Colin F. Camerer

  limits for economic analysis (market equilibrium, IO, public fi nance, etc.).

  A diff erent approach is to defi ne behavioral economics more generally, as

  simply being open- minded about what variables are likely to infl uence eco-

  nomic choices.

  This open- mindedness can be defi ned by listing neighboring social

  sciences that are likely to be the most fruitful source of explanatory variables.

  These include psychology, sociology (e.g., norms), anthropology (cultural

  variation in cognition), neuroscience, political science, and so forth. Call this

  the “behavioral economics trades with its neighbors” view.

  But the open- mindedness could also be characterized even more gener-

  ally, as an invitation to machine- learn how to predict economic outcomes

  from the largest possible feature set. In the “trades with its neighbors” view,

  features are constructs that are contributed by diff erent neighboring sciences.

  These could be loss aversion, identity, moral norms, in-group preference,

  inattention, habit, model- free reinforcement learning, individual polygenic

  scores, and so forth.

  But why stop there?

  In a general ML approach, predictive features could be—and should

  be— any variables that predict. (For policy purposes, variables that could

  be controlled by people, fi rms, and governments may be of special interest.)

  These variables can be measurable properties of choices, the set of choices,

  aff ordances and motor interactions during choosing, measures of atten-

  tion, psychophysiological measures of biological states, social infl uences,

  properties of individuals who are doing the choosing (SES, wealth, moods,

  personality, genes), and so forth. The more variables, the merrier.

  From this perspective, we can think about what sets of features are con-

  tributed by diff erent disciplines and theories. What features does textbook

  economic theory contribute? Constrained utility maximization in its most

  familiar and simple form points to only three kinds of variables—prices,

  information (which can inform utilities), and constraints.

  Most propositions in behavioral economics add some variables to this

  list of features, such as reference- dependence, context- dependence (menu

  eff ects), anchoring, limited attention, social preference, and so forth.

  Going beyond familiar theoretical constructs, the ML approach to behav-

  ioral economics specifi es a very long list of candidate variables (= features)

  and include all of them in an ML approach. This approach has two advan-

  tages: First, simple theories can be seen as bets that only a small number of

  features will predict well; that is, some eff ects (such as prices) are hypoth-

  esized to be fi rst- order in magnitude. Second, if longer lists of features pre-

  dict better than a short list of theory- specifi ed features, then that fi nding

  establishes a plausible upper bound on how much potential predictability

  is left to understand. The results are also likely to create raw material for

  theory to fi gure out how to consolidate the additional predictive power into

  crystallized theory (see also Kleinberg, Liang, and Mullainathan 2015).

  Artifi cial Intelligence and Behavioral Economics 589

  If behavioral economics is recast as open- mindedness about what vari-

  ables might predict, then ML is an ideal way to do behavioral economics

  because it can make use of a wide set of variables and select which ones

  predict. I will illustrate it with some examples.

  Bargaining. There is a long history of bargaining experiments trying to

  predict what bargaining outcomes (and disagreement rates) will result from

  structural variables using game- theoretic methods. In the 1980s there was

  a sharp turn in experimental work toward noncooperative approaches in

  which the communication and structure of bargaining was carefully struc-

  tured (e.g., Roth 1995 and Camerer 2003 for reviews). In these experiments

  the possible sequence of off ers in the bargaining are heavily constrained

  and no communication is allowed (beyond the off ers themselves). This

  shift to highly structured paradigms occurred because game theory, at the

  time, delivered sharp, nonobvious new predictions about what outcomes

  might result depending on the structural parameters—particularly, costs

  of delay, time horizon, the exogenous order of off ers and acceptance, and

  available outside options (payoff s upon disagreement). Given the diffi

  culty

  of measuring or controlling these structural variables in most fi eld settings,

  experiments provided a natural way to test these structured- bargaining

  theories.1

  Early experiments made it clear that concerns for fairness or outcomes

  of others infl uenced utility, and the planning ahead assumed in subgame

  perfect theories is limited and cognitively unnatural (Camerer et al. 1994;

  Johnson et al. 2002; Binmore et al. 2002). Experimental economists became

  wrapped up in understanding the nature of apparent social preferences and

  limited planning in structured bargaining.

  However, most natural bargaining is not governed by rules about structure

  as simple as those theories, and experiments became focused from 1985 to

  2000 and beyond. Natural bargaining is typically “semi- structured”—that

  is, there is a hard deadline and protocol for what constitutes an agreement,

  and otherwise there are no restrictions on which party can make what off ers

  at what time, including the use of natural language, face- to-face meetings

  or use of agents, and so on.

  The revival of experimental study of unstructured bargaining is a good

  idea for three reasons (see also Karagözog˘lu, forthcoming). First, there are

  now a lot more ways to measure what happens during bargaining in labora-

  tory conditions (and probably in fi eld settings as well). Second, the large

  number of features that can now be generated are ideal inputs for ML to

  predict bargaining outcomes. Third, even when bargaining is unstructured

  it is possible to produce bold, nonobvious precise predictions (thanks to the
>
  revelation principle). As we will see, ML can then test whether the features

  1. Examples include Binmore, Shaked, and Sutton (1985, 1989); Neelin, Sonnenschein, and Spiegel (1988); Camerer et al. (1994); and Binmore et al. (2002).

  590 Colin F. Camerer

  Fig. 24.1 A, initial off er screen (for informed player I, white bar); B, example cursor locations after three seconds (indicating amount off ered by I, white, or demanded by U, dark gray); C, cursor bars match which indicates an off er, consummated at six seconds; D, feedback screen for player I. Player U also receives feedback about pie size and profi t if a trade was made (otherwise the profi t is zero).

  predicted by game theory to aff ect outcomes actually do, and how much

  predictive power other features add (if any).

  These three properties are illustrated by experiments of Camerer, Nave,

  and Smith (2017).2 Two players bargain over how to divide an amount of

  money worth $1– $6 (in integer values). One informed ( I ) player knows the

  amount; the other, uninformed ( U ) player, doesn’t know the amount. They

  are bargaining over how much the uninformed U player will get. But both

  players know that I knows the amount.

  They bargain over ten seconds by moving cursors on a bargaining number

  line (fi gure 24.1). The data created in each trial is a time series of cursor loca-

  tions, which are a series of step functions coming from a low off er to higher

  ones (representing increases in off ers from I ) and from higher demands to

  lower ones (representing decreasing demands from U ).

  Suppose we are trying to predict whether there will be an agreement or

  not based on all variables that can be observed. From a theoretical point

  of view, effi

  cient bargaining based on revelation principle analysis predicts

  an exact rate of disagreement for each of the amounts $1– 6, based only on

  the diff erent amounts available. Remarkably, this prediction is process- free.

  2. This paradigm builds on seminal work on semistructured bargaining by Forsythe, Ken-

  nan, and Sopher (1991).

  Artifi cial Intelligence and Behavioral Economics 591

  Fig. 24.2 ROC curves showing combinations of false and true positive rates in pre-

  dicting bargaining disagreements

  Notes: Improved forecasting is represented by curves moving to the upper left. The combination of process (cursor location features) and “pie” (amount) data are a clear improvement over either type of data alone.

  However, from an ML point of view there are lots of features represent-

  ing what the players are doing that could add predictive power (besides the

  process- free prediction based on the amount at stake). Both cursor locations

  are recorded every twenty- fi ve msec. The time series of cursor locations is

  associated with a huge number of features—how far apart the cursors are,

  the time since last concession (= cursor movement), size of last concession,

  interactions between concession amounts and times, and so forth.

  Figure 24.2 shows an ROC curve indicating test- set accuracy in predicting

  whether a bargaining trial ends in a disagreement (= 1) or not. The ROC

  curves sketch out combinations of true positive rates, P(disagree|predict

  disagree) and false positive rates P(agree|predict disagree). An improved

  ROC curve moves up and to the left, refl ecting more true positives and fewer

  false positives. As is evident, predicting from process data only is about as

  accurate as using just the amount (“pie”) sizes (the ROC curves with black

  circle and empty square markers). Using both types of data improves predic-

  tion substantially (curve with empty circle markers).

  Machine learning is able to fi nd predictive value in details of how the

  bargaining occurs (beyond the simple, and very good, prediction based

  only on the amount being bargained over). Of course, this discovery is the

  592 Colin F. Camerer

  beginning of the next step for behavioral economics. It raises questions that

  include: What variables predict? How do emotions,3 face- to-face commu-

  nication, and biological measures (including whole- brain imaging)4 infl u-

  ence bargaining? Do people consciously understand why those variables are

  important? Can ML methods capture the eff ects of motivated cognition in

  unstructured bargaining, when people can self- servingly disagree about case

  facts?5 Can people constrain expression of variables that hurt their bargain-

  ing power? Can mechanisms be designed that record these variables and

  then create effi

  cient mediation, into which people will voluntarily participate

  (capturing all gains from trade)?6

  Risky Choice. Peysakhovich and Naecker (2017) use machine learning to

  analyze decisions between simple fi nancial risks. The set of risks are ran-

  domly generated triples ($ y, $ x, 0) with associated probabilities ( p _ x, p _ y, p _0). Subjects give a willingness- to-pay (WTP) for each gamble.

  The feature set is the fi ve probability and amount variables (excluding the

  $0 payoff ), quadratic terms for all fi ve, and all two- and three- way inter-

  actions among the linear and quadratic variables. For aggregate- level esti-

  mation this creates 5 + 5 + 45 + 120 = 175 variables.

  Machine learning predictions are derived from regularized regression

  with a linear penalty (LASSO) or squared penalty (ridge) for (absolute)

  coeffi

  cients. Participants were N = 315 MTurk subjects who each gave ten

  useable responses. The training set consists of 70 percent of the observa-

  tions, and 30 percent are held out as a test set.

  They also estimate predictive accuracy of a one- variable expected utility

  model (EU, with power utility) and a prospect theory (PT) model, which

  adds one additional parameter to allow nonlinear probability weighting

  (Tversky and Kahneman 1992) (with separate weights, not cumulative ones).

  For these models there are only one or two free parameters per person.7

  The aggregate data estimation uses the same set of parameters for all

  subjects. In this analysis, the test set accuracy (mean squared error) is almost

  exactly the same for PT and for both LASSO and ridge ML predictions, even

  though PT uses only two variables and the ML methods use 175 variables.

  Individual- level analysis, in which each subject has their own parameters

  has about half the mean squared error as the aggregate analysis. The PT and

  ridge ML are about equally accurate.

  The fact that PT and ML are equally accurate is a bit surprising because

  the ML method allows quite a lot of fl exibility in the space of possible

  3. Andrade and Ho (2009).

  4. Lohrenz et al. (2007) and Bhatt et al. (2010).

  5. See Babcock et al. (1995) and Babcock and Loewenstein (1997).

  6. See Krajbich et al. (2008) for a related example of using neural measures to enhance effi

  -

  ciency in public good production experiments.

  7. Note, however, that the ML feature set does not exactly nest the EU and PT forms. For example, a weighted combination of the linear outcome X and the quadratic term X 2 does not exactly equal the power function X.

  Artifi cial Intelligence and Behavioral Economics 593

  predictions. Indeed, the authors’ motivation was to use ML to show how

  a model with a huge amount of fl exibility could fi
t, possibly to provide a

  ceiling in achievable accuracy. If the ML predictions were more accurate

  than EU or PT, the gap would show how much improvement could be had

  by more complicated combinations of outcome and probability parameters.

  But the result, instead, shows that much busier models are not more accurate

  than the time- tested two- parameter form of PT, for this domain of choices.

  Limited Strategic Thinking. The concept of subgame perfection in game

  theory presumes that players look ahead in the future to what other players

  might do at future choice nodes (even choice nodes that are unlikely to be

  reached), in order to compute likely consequences of their current choices.

  This psychological presumption does have some predictive power in short,

  simple games. However, direct measures of attention (Camerer at al. 1994;

  Johnson et al. 2002) and inference from experiments (e.g., Binmore et al.

  2002) make it clear that players with limited experience do not look far

  ahead.

  More generally, in simultaneous games, there is now substantial evi-

  dence that even highly intelligent and educated subjects do not all process

  information in a way that leads to optimized choices given (Nash) “equi-

  librium” beliefs—that is, beliefs that accurately forecast what other players

  will do. More important, two general classes of theories have emerged that

  can account for deviations from optimized equilibrium theory. One class,

  quantal response equilibrium (QRE), are theories in which beliefs are sta-

  tistically accurate but noisy (e.g., Goeree, Holt, and Palfrey 2016). Another

  type of theory presumes that deviations from Nash equilibrium result from

  a cognitive hierarchy of levels of strategic thinking. In these theories there

  are levels of thinking, starting from nonstrategic thinking, based presumably

  on salient features of strategies (or, in the absence of distinctive salience,

  random choice). Higher- level thinkers build up a model of what lower-

  level thinkers do (e.g., Stahl and Wilson 1995; Camerer, Ho, and Chong

  2004; Crawford, Costa-Gomes, and Iriberri 2013). These models have been

  applied to hundreds of experimental games with some degree of imperfect

  cross- game generality, and to several fi eld settings.8

 

‹ Prev