Showing posts with label learning. Show all posts
Showing posts with label learning. Show all posts

Monday, October 24, 2022

Informationally Simple Incentives by Simon Gleyze and Agathe Pernoud

 Agathe Pernoud is on the Economics job market from Stanford this year, and is interested in the properties of information in environments in which agents may need to learn their own preferences.

Here are two papers that advance the theory of those situations, and expand on the fragility of 'dominant strategies' as the strategy space is enlarged.

Informationally Simple Incentives by Simon Gleyze and Agathe Pernoud, Journal of Political Economy, forthcoming.

Abstract: We consider a mechanism design setting in which agents can acquire costly information on their preferences as well as others’. A mechanism is informationally simple if agents have no incentive to learn about others’ preferences. This property is of interest for two reasons: First, it is a necessary condition for the existence of dominant strategy equilibria in the extended game.  Second, this endogenizes an “independent private value” property of the interim information structure. We show that, generically, a mechanism is informationally simple if and only if it satisfies a separability condition which rules out most economically meaningful mechanisms."


See also Agathe's job market paper:

How Competition Shapes Information in Auctions by Simon Gleyze and Agathe Pernoud

We consider auctions where buyers can acquire costly information about their valuations and those of others, and investigate how competition between buyers shapes their learning incentives. In equilibrium, buyers find it cost-efficient to acquire some information about their competitors so as to only learn their valuations when they have a fair chance of winning. We show that such learning incentives make competition between buyers less effective: losing buyers often fail to learn their valuations precisely and, as a result, compete less aggressively for the good. This depresses revenue, which remains bounded away from the expected second-highest valuation even when information costs are small. It also undermines price discovery. Finally, we examine the implications for auction design. First, setting an optimal reserve price is more valuable than attracting an extra buyer. Second, the seller can incentivize buyers to learn their valuations, hence restoring effective competition, by maintaining uncertainty over the set of auction participants.


Wednesday, August 24, 2022

Learning and competition in the lab, in France, and in India

 Three NBER working papers this week particularly caught my eye: a lab experiment, a natural experiment, and a field experiment.

The first is a reminder of why simple reinforcement learning models have as much predictive power as they do. It's an experiment that shows that even when others' experience is made clearly visible, there's a tendency to rely on 'own experience'.

Not Learning from Others by John J. Conlon, Malavika Mani, Gautam Rao, Matthew W. Ridley & Frank Schilbach  WORKING PAPER 30378 DOI 10.3386/w30378 August 2022

Abstract: We provide evidence of a powerful barrier to social learning: people are much less sensitive to information others discover compared to equally-relevant information they discover themselves. In a series of incentivized lab experiments, we ask participants to guess the color composition of balls in an urn after drawing balls with replacement. Participants' guesses are substantially less sensitive to draws made by another player compared to draws made themselves. This result holds when others' signals must be learned through discussion, when they are perfectly communicated by the experimenter, and even when participants see their teammate drawing balls from the urn with their own eyes. We find a crucial role for taking some action to generate one's `own' information, and rule out distrust, confusion, errors in probabilistic thinking, up-front inattention and imperfect recall as channels.

******

The second is a careful study of affirmative action for women in French chess tournaments: a requirement that teams include a woman had many effects, including improvement in the quality of play by French women.

Trickle-Down Effects of Affirmative Action: A Case Study in France by José De Sousa & Muriel Niederle, WORKING PAPER 30367 DOI 10.3386/w30367 August 2022

Abstract: "The introduction of a quota in the French chess Club Championship in 1990, an activity many players engage in next to playing in individual tournaments, provides a quite unique environment to study its effects on three levels. We find that women selected by the quota improve their performance. We show large spillover and trickle-down effects: There are more and better qualified women. International comparisons confirm that the results are unique to France and that there are no substantial adverse effects on French male players. We discuss the properties of this quota and how to implement it in other environments."

The concluding paragraph:

"We speculate that one reason for the success of the French chess quota was due to the fact that it was an “output” rather than a “pure representation” quota. At least one ninth of the performance of teams in the Club Championship was determined by the performance of female players. Such an “output” based quota provides organization with different incentives than a pure representation quota does. We use economic departments to discuss the different gender quotas and how each of them might be implemented. We hope that future work will provide theoretical properties of various quotas as well as find other areas where output quotas are already, or could be, implemented."

*********

The third is about the difficulty of inducing competition in close quarters.

Does the Invisible Hand Efficiently Guide Entry and Exit? Evidence from a Vegetable Market Experiment in India by Abhijit Banerjee, Greg Fischer, Dean Karlan, Matt Lowe & Benjamin N. Roth, WORKING PAPER 30360 DOI 10.3386/w30360, August 2022

Abstract: "What accounts for the ubiquity of small vendors operating side-by-side in the urban centers of developing countries? Why don’t competitive forces drive some vendors out of the market? We ran an experiment in Kolkata vegetable markets in which we induced (via subsidizing) some vendors to sell additional produce. The vendors earned higher profits, even when excluding the value of the subsidy. Nevertheless, after the subsidies ended vendors largely stopped selling the additional produce. Our results are consistent with collusion and inertial business practices suppressing competition and efficient market exit."


Friday, December 24, 2021

Costly information gathering to form preferences in school choice

 Here's a model suggesting that people for whom it is more costly to gather information about school quality will do less well in preference based school choice.

Inattention and Inequity in School Matching, by Stefan F. Bucher & Andrew Caplin, NBER WORKING PAPER 29586, DOI 10.3386/w29586

Abstract: The attractive properties of the Deferred Acceptance (DA) algorithm rest on the assumption of perfect information. Yet field studies of school matching show that information is imperfect, particularly for disadvantaged students. We model costly strategic learning when schools are ex ante symmetric, agree on their ranking of students, and learning is rationally inattentive. Our analytic solution quantifies how each student’s rank, learning costs and prior beliefs interact to determine their gross and net welfare as well as the extent and form of mistakes they make. In line with the evidence, we find that lower-ranked students are affected disproportionately more by information costs, generally suffering a larger welfare loss than higher-ranked students. Interactions between mechanism design, inattention and inequity are thus of first order importance.

**********

"The challenge faced by matching models with endogenous information is that students face three sources of uncertainty: signal-based, deriving from uncertainty about what information their learning strategy will produce; strategic, deriving from uncertainty about others’ submissions and thus the resulting matching outcome; and value-based, referring to the remaining uncertainty about the student’s valuation of their tch.

We introduce a tractable model of strategically rational inattention in a matching market that parsimoniously captures this complexity. To focus on the interplay with inequity we assume that schools agree on their ranking of students. For analytic tractability we assume that schools are ex ante symmetric (exchangeable) and that learning is rationally inattentive (Sims, 2003; Caplin and Dean, 2015; Matejka and McKay, 2015). While our symmetry assumption implies that schools are ex ante identical, it does not require that students’ valuations are independent across schools so that information on a school can update beliefs about others.

...

"A central finding is that DA exacerbates inequity. Lower-ranked students attain a lower fraction of their net welfare surplus under full information than do higher-ranked students, even if they have the same costs of learning. This is because lower-ranked students face greater uncertainty about the outcome resulting from any submission, which disperses and often dilutes their incentive to acquire information.

...

"The fact that lower-ranking students are more likely to be matched with a school further down their list results in very unequal learning incentives..."


Tuesday, November 30, 2021

Interview with Ido Erev

 Here's a short interview with Ido Erev, the great behavioral scientist at the Technion, from whom I learned a lot about learning:

Interview with Ido Erev

Here is his closing comment:

"When I started  studying the effect of experience, in the 90s, I asked several famous  researchers why they have stopped studying this effect. Here are some of  the answers that affected me the most (as I remember them): Duncan Luce: Now that I am old, I am more interested in my own mistakes.  In particular, I try to understand why I exhibit the Allais paradox.  Herb Simon: I got more reinforcements from studying bounded rationality.  Amos Tversky: It is clear that if you hit subjects with a 5Kg “feedback hammer” they will learn to be rational. I  want to study what people learn before they arrive at the lab. Alvin E.  Roth: There is no good answer, let's study it."

Wednesday, February 13, 2019

Market design through machine learning: David Parkes

If I were in Boston I'd go to hear David Parkes speak today about

Optimal Economic Design through Deep Learning

Abstract: Designing an auction that maximizes expected revenue is a major open problem in economics. Despite significant effort, only the single-item case is fully understood. We ask whether the tools of deep learning can be used to make progress. We show that multi-layer neural networks can learn essentially optimal auction designs for the few problems that have been solved analytically, and can be used to design auctions for poorly understood problems, including settings with multiple items and budget constraints. I will also overview applications to other problems of optimal economic design, and discuss the broader implications of this work. Joint work with Paul Duetting (London School of Economics), Zhe Feng (Harvard University), Noah Golowich (Harvard University), Harikrishna Narasimhan (Harvard -> Google), and Sai Srivatsa (Harvard University). Working paper: https://arxiv.org/abs/1706.03459

Tuesday, March 29, 2016

Whither Game Theory? by Fudenberg and Levine (we need to learn more about learning)


Whither Game Theory? Drew Fudenberg David K. Levine, January 31, 2016

Abstract: We examine the state of game theory. Many important questions have been answered, and
game theoretic methods are now central to much economic investigation. We suggest areas where
further advances are important, and argue that models of learning and of social preferences
provide promising routes for improving and widening game theory’s predictive power, while
preserving the sucesses of existing theory where it works well. We emphasize in particular the
need for better understanding of the speed with which learning takes place, and of the evolution
of social norms.

Friday, July 3, 2015

Arrow Lecture in Jerusalem by Drew Fudenberg - Learning and Equilibrium in Games (video)

Drew begins his general-audience lecture by saying "I can't imagine anyone I would rather give a talk for than Ken Arrow." He then continues with a brief history of game theory.

Tuesday, July 15, 2014

Learning and adaptation in the social sciences, in the PNAS

The papers from the Learning and adaptation in the social sciences: NAS Sackler conference, January 10-11 (including my paper with Ido Erev) are now online in the 'early edition' of the Proceedings of the National Academy of Sciences (PNAS):


Here's the Abstract of that last paper, if you got this far:)

Abstract

The rationality assumption that underlies mainstream economic theory has proved to be a useful approximation, despite the fact that systematic violations to its predictions can be found. That is, the assumption of rational behavior is useful in understanding the ways in which many successful economic institutions function, although it is also true that actual human behavior falls systematically short of perfect rationality. We consider a possible explanation of this apparent inconsistency, suggesting that mechanisms that rest on the rationality assumption are likely to be successful when they create an environment in which the behavior they try to facilitate leads to the best payoff for all agents on average, and most of the time. Review of basic learning research suggests that, under these conditions, people quickly learn to maximize expected return. This review also shows that there are many situations in which experience does not increase maximization. In many cases, experience leads people to underweight rare events. In addition, the current paper suggests that it is convenient to distinguish between two behavioral approaches to improve economic analyses. The first, and more conventional approach among behavioral economists and psychologists interested in judgment and decision making, highlights violations of the rational model and proposes descriptive models that capture these violations. The second approach studies human learning to clarify the conditions under which people quickly learn to maximize expected return. The current review highlights one set of conditions of this type and shows how the understanding of these conditions can facilitate market design.