Computational environment design

The Internet has evolved into a platform on which large numbers of individuals take action and join in collaborations via crowdsourcing, social media, and electronic commerce. When designing social and economic systems on the Internet, a key challenge is understanding how to promote particular desired behaviors and outcomes. I call this problem computational environment design. Notable abilities afforded by the Internet, such as the ability to recruit large numbers of individuals to join problem-solving efforts via crowdsourcing and social media, and the ability to engage in a data-driven iterative design process, are creating new opportunities and inspiring new methods for computational environment design. This dissertation focuses on these abilities and proposes an approach for arriving at effective designs by reasoning and learning about characteristics of participants and how these characteristics interact with a system's design to influence behavior. The dissertation consists of two major components. The first component focuses on designing crowdsourcing and human computation systems that leverage a crowd to solve complex problems that require effective coordination among participants or the recruitment of individuals with relevant expertise. I show how reasoning about crowd abilities and limitations can lead to designs that make crowdsourcing complex tasks feasible, effective, and efficient. The solutions introduce new design patterns and methods for human computation and crowdsourcing; notable contributions include a crowdware design for tackling human computation tasks with global constraints, and incentive mechanisms for task routing that harness people's expertise and social expertise by engaging them in both problem solving and routing. The second component focuses on understanding how to design effective environments automatically. I introduce a general active, indirect elicitation framework for automated environment design that learns relevant characteristics of participants based on observations of their behavior and optimizes designs based on learned models. Theoretical contributions include developing an active, indirect elicitation algorithm for a sequential decision-making setting that is guaranteed to discover effective designs after few interactions. Practical contributions include applications of the active, indirect elicitation framework to crowdsourcing. Specifically, I demonstrate how to automatically design tasks and synthesize workflows when optimizing for desired objectives given resource constraints.

互联网已经演变成一个平台，大量个体通过众包、社交媒体和电子商务采取行动并参与合作。在互联网上设计社会和经济系统时，一个关键挑战是了解如何促进特定的期望行为和结果。我将这个问题称为计算环境设计。 互联网提供的显著能力，例如通过众包和社交媒体招募大量个体参与解决问题的能力，以及进行数据驱动的迭代设计过程的能力，正在为计算环境设计创造新的机遇并激发新的方法。本论文聚焦于这些能力，并提出一种通过对参与者的特征进行推理和学习以及这些特征如何与系统设计相互作用以影响行为来实现有效设计的方法。 本论文由两个主要部分组成。第一部分侧重于设计众包和人类计算系统，这些系统利用群体来解决需要参与者之间有效协调或招募具有相关专业知识的个体的复杂问题。我展示了对群体能力和局限性的推理如何能够导致使众包复杂任务可行、有效和高效的设计。这些解决方案为人类计算和众包引入了新的设计模式和方法；显著的贡献包括一种用于处理具有全局约束的人类计算任务的群体软件设计，以及通过让人们参与解决问题和任务分配来利用人们的专业知识和社会专业知识的任务分配激励机制。 第二部分侧重于了解如何自动设计有效的环境。我引入了一个用于自动环境设计的通用主动、间接诱导框架，该框架根据对参与者行为的观察来学习参与者的相关特征，并根据学习到的模型优化设计。理论贡献包括为顺序决策环境开发一种主动、间接诱导算法，该算法在经过少量交互后保证能发现有效的设计。实践贡献包括主动、间接诱导框架在众包中的应用。具体来说，我演示了在给定资源约束的情况下，为优化期望目标如何自动设计任务和合成工作流程。