Predictive models of human behavior in strategic settings

拟议的项目将开发通用模型，准确预测真实的人在战略环境中的行为。这样的模型很重要，因为这样的设置的设计--例如，谷歌为运行搜索关键词旁边的广告时段拍卖而建立的规则--关键取决于参与者的行为模型。具体地说，设计者选择产生最佳预测行为的规则，例如，优化设计者的收入或所有参与者的福利。目前，参与者通常被建模为行为“经济理性”。当所有参与者都是大公司时，这一假设可以说是合理的。然而，最近的大量实验证据表明，它往往不适用于人类个体。因此，据说是最优的设计很可能在实践中表现得不太理想。如果人们只是表现得不确定，这可能就是故事的结局。然而，人们不仅仅是非理性的--他们也是可以预见的非理性的。在过去十年左右的时间里，行为博弈论者对人们偏离理性的各种系统方式进行了分类。然而，这篇文献并没有告诉我们如何模拟真实的人在任意的战略环境中的行为。我的长期研究目标是生成完整的人类行为预测模型，并使用机器学习中最先进的方法来验证这些模型。这项提案中描述的工作将遵循四个相互关联的工作思路：(1)对突出度在人类战略发挥中的作用进行建模；(2)利用最近的深度学习思想来自动发现特征；(3)开发预测一个人的战略推理深度的模型；以及(4)研究如何在我们的模型预测的行为类型下，优化谷歌的广告拍卖设计，以获得更好的性能。总体而言，更准确的竞标者行为预测模型对改善谷歌关键字拍卖的运营有直接影响，而关键字拍卖是加拿大各地小企业的关键收入来源，甚至对各种行业的市场优化也有直接影响。

The proposed project will develop general models that accurately predict how real people behave in strategic settings. Such models are important because the design of such settings--e.g., the rules that Google establishes to run its auctions for advertising slots alongside search keywords--depends crucially on models of participant behavior. Specifically, a designer chooses rules that yield the best predicted behavior, e.g., optimizing the designer's revenue or the welfare of all participants. Currently, participants are usually modeled as behaving "economically rationally". This assumption is arguably reasonable when all participants are large corporations. However, a wealth of recent experimental evidence shows that it very often fails to hold for individual human beings. Thus, designs said to be optimal are likely to perform suboptimally in practice. If people simply behaved nondeterministically, this might be the end of the story. However, people are not just irrational--they are predictably irrational. Over the last decade or so, behavioral game theorists have catalogued a wide range of systematic ways in which people deviate from rationality. However, this literature doesn't quite tell us how to model how real people behave in arbitrary strategic settings. My long-term research goal is to produce full predictive models of human behavior, and to validate these models using state-of-the-art methods from machine learning. The work described in this proposal will pursue four interrelated threads of work: (1) modeling the effect of salience in human strategic play; (2) leveraging the recent idea of deep learning to automate feature discovery; (3) developing models that predict a person's depth of strategic reasoning; and (4) investigating ways that Google's ad auction design could be optimized for better performance under the sort of behavior predicted by our models. Overall, more accurate predictive models of bidder behavior have immediate implications for improving the operation of Google's keyword auctions--a key revenue driver for small businesses across Canada--and indeed for the optimization of markets across a wide variety of industries.