Person:

Parkes, David

A central challenge in human computation is in understanding how to design task environments that effectively attract participants and coordinate the problem solving process. In this paper, we consider a common problem that requesters face on Amazon Mechanical Turk: how should a task be designed so as to induce good output from workers? In posting a task, a requester decides how to break down the task into unit tasks, how much to pay for each unit task, and how many workers to assign to a unit task. These design decisions affect the rate at which workers complete unit tasks, as well as the quality of the work that results. Using image labeling as an example task, we consider the problem of designing the task to maximize the number of quality tags received within given time and budget constraints. We consider two different measures of work quality, and construct models for predicting the rate and quality of work based on observations of output to various designs. Preliminary results show that simple models can accurately predict the quality of output per unit task, but are less accurate in predicting the rate at which unit tasks complete. At a fixed rate of pay, our models generate different designs depending on the quality metric, and optimized designs obtain significantly more quality tags than baseline comparisons.

We present a direction of research on principles and methods that can enable general problem solving via human computation systems. A key challenge in human computation is the effective and efficient coordination of problem solving. While simple tasks may be easy to partition across individuals, more complex tasks highlight challenges and opportunities for more sophisticated coordination and optimization, leveraging such core notions as problem decomposition, subproblem routing and solution, and the recomposition of solved subproblems into solutions. We discuss the interplay between algorithmic paradigms and human abilities,and illustrate through examples how members of a crowd can play diverse roles in an organized problem-solving process, serving not only as "data oracles" at the endpoints of computation, but also as modules for decomposing problems, controlling the algorithmic progression, and performing human program synthesis.

With the proliferation of online labor markets and other social computing platforms, online experiments have become a low-cost and scalable way to empirically test hypotheses and mechanisms in both human computation and social science. Yet, despite the potential in designing more powerful and expressive online experiments using multiple subjects, researchers still face many technical and logistical difficulties. We see synchronous and longitudinal experiments involving real-time interaction between participants as a dual-use paradigm for both human computation and social science, and present TurkServer, a platform that facilitates these types of experiments on Amazon Mechanical Turk. Our work has the potential to make more fruitful online experiments accessible to researchers in many different fields.

By coordinating efforts from humans and machines, human computation systems can solve problems that machines cannot tackle alone. A general challenge is to design efficient human computation algorithms or workflows with which to coordinate the work of the crowd. We introduce a method for automated workflow synthesis aimed at ideally harnessing human efforts by learning about the crowd's performance on tasks and synthesizing an optimal workflow for solving a problem. We present experimental results for human sorting tasks, which demonstrate both the benefit of understanding and optimizing the structure of workflows based on observations. Results also demonstrate the benefits of using value of information to guide experiments for identifying efficient workflows with fewer experiments.

Many situations arise in which an interested party's utility is dependent on the actions of an agent; e.g., a teacher is interested in a student learning effectively and a firm is interested in a consumer's behavior. We consider an environment in which the interested party can provide incentives to affect the agent's actions but cannot otherwise enforce actions. In value-based policy teaching, we situate this within the framework of sequential decision tasks modeled by Markov Decision Processes, and seek to associate limited rewards with states that induce the agent to follow a policy that maximizes the total expected value of the interested party. We show value-based policy teaching is NP-hard and provide a mixed integer program formulation. Focusing in particular on environments in which the agent's reward is unknown to the interested party, we provide a method for active indirect elicitation wherein the agent's reward function is inferred from observations about its response to incentives. Experimental results suggest that we can generally find the optimal incentive provision in a small number of elicitation rounds.

Policy teaching considers a Markov Decision Process setting in which an interested party aims to influence an agent's decisions by providing limited incentives. In this paper, we consider the specific objective of inducing a pre-specified desired policy. We examine both the case in which the agent's reward function is known and unknown to the interested party, presenting a linear program for the former case and formulating an active, indirect elicitation method for the latter. We provide conditions for logarithmic convergence, and present a polynomial time algorithm that ensures logarithmic convergence with arbitrarily high probability. We also offer practical elicitation heuristics that can be formulated as linear programs, and demonstrate their effectiveness on a policy teaching problem in a simulated ad-network setting. We extend our methods to handle partial observations and partial target policies, and provide a game-theoretic interpretation of our methods for handling strategic agents.

Activity rules have emerged in recent years as an important aspect of practical auction design. The role of an activity rule in an iterative auction is to suppress strategic behavior by bidders and promote simple, continual, meaningful bidding and thus, price discovery. These rules find application in the design of iterative combinatorial auctions for real world scenarios, for example in spectrum auctions, in airline landing slot auctions, and in procurement auctions. We introduce the notion of strong activity rules, which allow simple, consistent bidding strategies while precluding all behaviors that cannot be rationalized in this way. We design such a rule for auctions with budget-constrained bidders, i.e., bidders with valuations for resources that are greater than their ability to pay. Such bidders are of practical importance in many market environments, and hindered from bidding in a simple and consistent way by the commonly used revealed-preference activity rule, which is too strong in such an environment. We consider issues of complexity, and provide two useful forms of information feedback to guide bidders in meeting strong activity rules. As a special case, we derive a strong activity rule for non-budget-constrained bidders. The ultimate choice of activity rule must depend, in part, on beliefs about the types of bidders likely to participate in an auction event because one cannot have a rule that is simultaneously strong for both budget-constrained bidders and quasi-linear bidders.

The problem of environment design considers a setting in which an interested party aims to influence an agent's decisions by making limited changes to the agent's environment. Zhang and Parkes [2008] first introduced the environment design concept for a specific problem in the Markov Decision Process setting. In this paper, we present a general framework for the formulation and solution of environment design problems with one agent. We consider both the case in which the agent's local decision model is known and partially unknown to the interested party, and illustrate the framework and results on a linear programming setting. For the latter problem, we formulate an active, indirect elicitation method and provide conditions for convergence and logarithmic convergence. We relate to the problem of inverse optimization and also offer a game-theoretic interpretation of our methods.

Many situations arise in which an interested party wishes to affect the decisions of an agent; e.g., a teacher that seeks to promote particular study habits, a Web 2.0 site that seeks to encourage users to contribute content, or an online retailer that seeks to encourage consumers to write reviews. In the problem of environment design, one assumes an interested party who is able to alter limited aspects of the environment for the purpose of promoting desirable behaviors. A critical aspect of environment design is understanding preferences, but by assumption direct queries are unavailable. We work in the inverse reinforcement learning framework, adopting here the idea of active indirect preference elicitation to learn the reward function of the agent by observing behavior in response to incentives. We show that the process is convergent and obtain desirable bounds on the number of elicitation rounds. We briefly discuss generalizations of the elicitation method to other forms of environment design, e.g., modifying the state space, transition model, and available actions.

We consider a setting in which a principal seeks to induce an adaptive agent to select a target action by providing incentives on one or more actions. The agent maintains a belief about the value for each action—which may update based on experience—and selects at each time step the action with the maximal sum of value and associated incentive. The principal observes the agent’s selection, but has no information about the agent’s current beliefs or belief update process. For inducing the target action as soon as possible, or as often as possible over a fixed time period, it is optimal for a principal with a per-period budget to assign the budget to the target action and wait for the agent to want to make that choice. But with an across-period budget, no algorithm can provide good performance on all instances without knowledge of the agent’s update process, except in the particular case in which the goal is to induce the agent to select the target action once. We demonstrate ways to overcome this strong negative result with knowledge about the agent’s beliefs, by providing a tractable algorithm for solving the offline problem when the principal has perfect knowledge, and an analytical solution for an instance of the problem in which partial knowledge is available.