The tentative format for the initial talks is for an intern to talk 10 minutes about his/her previous work and 10-15 minutes about the plan for the summer. The summer plan should be detailed and specific.

A non-intern IDL member should talk about current work for 20 minutes.

The end-of-summer talks should be an intern talking for 20 minutes about the results of summer work.

Thursdays from 10-12.

DateTimeLocation
6/2910AM-12PM Kappa

1. Dave Levin

  • Previous Work: Patience and Punishment: Incentive Mechanisms in Selfish Networks
  • Summer Plan: Cold-Start Multicast in Tycoon
  • Abstract: We address an all-too-common scalability issue when running an experiment on a distributed set of providing nodes, P. Often, the experimenter accesses a single machine, s, but needs to distribute messages (e.g., small control messages or large data files) to the entire set P. Though multicast is an obvious, efficient solution to this problem, it is not used in practice. This is because of the following fundamental difference: in standard multicast, each recipient opts in to the system, but in this problem of cold-start multicast, the source s must convince each member of P to participate and to create the multicast structure on the fly. We will investigate application-layer multicast in this new light, and explore methods that offer various combinations of latency to send to all nodes in P, resilience to node and link failure, and overhead.

1. Thomas Sandholm

  • Title: Performance Prediction and Reservations in Computational Grid Markets
  • Previous Work: QoS and Price Prediction in Competitive Computational Markets
  • Summer Plan: Lightweight Reservations for the Tycoon Grid Broker
  • Abstract: Computational Markets allow fine-grained and adaptive control over delivered service levels. However, agility comes at the cost of predictability, making it hard for users to know how much money to spend to get the required service level with a certain likelihood. We have investigated and implemented a suite of prediction techniques that could help guide users to budget their Grid jobs more efficiently. Initial results have also been obtained from studying PlanetLab usage logs. In the future we hope to utilize these techniques in various Tycoon brokers to deliver service levels more reliably.

1. Scott Golder

  • Title: Social Information Organization: Tags and Other Personal Metadata
  • Abstract: We often rely on experts to organize things for us. Library books are organized by expert cataloguers, and reference books themselves are authored by domain experts. What happens when there are too many items for any expert to categorize, or more topics than there are experts? Social methods of information organization rely on the collective knowledge and action of large groups of contributors to do the organizing. In this talk, I will give an overview of this growing practice, incluing my own recent work in tagging for document and photo management, as well as other examples like Wikipedia and digg. I will also discuss a current piece of work, a laboratory experiment on imitation and influence in tagging, and the light the results can shed on information cascades and efficiency of social IO systems.
DateTimeLocation
7/610AM-12PM Tioga

1. Alvin AuYoung

  • Previous Work: Market-based resource allocation
  • Summer Plan: The economics of sharing: a case study of service deployment in PlanetLab
  • Abstract: Despite the potential gains of using market for resource allocation, very few market-based techniques have been adopted to perform allocation in shared computing infrastructures. The lack of adoption can be partially attributed to skepticism from system users and operators, who do not fully understand the tradeoffs involved in introducing a market. The goal of this work is construct a system X to demonstrate the tradeoffs associated with a real deploying of a market-based resource allocation system within a live shared wide-area infrastructure. What makes this study different from previous studies is that we show the benefit of deploying this system without using artificial utility functions, and without assuming full participation from end users. Since utility is difficult to derive from existing workloads, we quantify the benefit of a market from the standpoint of having load-balanced resources. It has been argued that in platforms with highly variable resource availability and demand, intelligent service placement or migration is necessary in order to allow the platform to maximize the value it delivers to applications. If a market is introduced, the resource prices it generates can provide implicit load signals to users and facilitate decentralized load-balancing in the system. This process allows the wide-area platform to achieve decentralized “smart placement” without the overhead of service migration [Oppenheimer06]. Using this metric provides a way of quantifying the benefits of the market without relying on synthetically-generated utility functions in our argument. Secondly, to address concerns of the additional burden it places on users, we analyze the effect of the market on users who do not participate (i.e. applications that require no changes from before the market structure was introduced), and for users who do. Moreover, we compare variation on the market structure, such as proportional-share and reservation-based resource guarantees.

1. Travis Kriplean

  • Previous Work:
  • Summer Plan: Communal Structuring in Wikipedia
  • Abstract:

This project seeks to identify the connection between the normative cooperative aspect of Wikipedia and the actual product that results. We explore this issue by identifying textual cues that reflect Wikipedian norms and track them across different categories of articles, exogenous events (e.g. Seigenthaler biography), and reflexive tagging of articles (e.g. "this article needs to be cleaned up"). Future work might use the categories and criteria that result to help categorize content and alert to unwanted communal dynamics or encourage desired interaction styles.

1. Fang Wu

  • Title:
  • Abstract:
DateTimeLocation
7/1310AM-12PM Tioga

1. Josep Pujol

  • Previous Work: Structure in Artificial Societies.

This will be a brief summary of the work conducted in my Ph.D. on Artificial Intelligence. How the structure of interactions between agents can be used to infer knowledge as well as to gain insight into the emergence of collective behaviors (in particular, conventions). Moreover, the formation of certain network structures is addressed by proposing a local interaction model grounded by sociologically plausible assumptions.

  • Summer Plan: Study of the structure of interactions in social processes.

1) First, we want to analyze the empirical data on the "reputation experiment" at the individual level. By analyzing the pattern of interactions between seller and buyers as well as the evolution of theses patterns we hope to find interesting regularities in the individual behavior. Thanks to the empirical data we can study how real people react against defection, retaliation or how individual reputation induces a competitive advantage. 2) Second, we intend to extend the work on decision insurance in organizations by taking into account the dynamics of opinion formation. To deter the temptation of free-riding introduced by the insurance, some sort of social control is required. The people better informed to make that decision (whether he is free-riding or not) are those who are close (social network) to the subject being monitored. However, social networks may sometimes have negative effects on the quality of decisions because people mainly seek (and get) from their ties support for what they want to do anyway, but do not get the necessary critical feedback to make better decisions.

1. Ali Ghodsi

  • Previous Work/Summer Plan: DHTs and the survival of unpopular items
  • Abstract:

This talk summarizes my past three years and my coming three months, focusing on the past. In my past I have designed and developed a structured overlay called DKS, which provides a Distributed Hash Table (DHT). First, a brief tutorial on DHTs is given. Thereafter, the distinguishing features of DKS are presented. These include lookup consistency guarantees, a simple replication scheme, efficient topology maintenance, and efficient overlay multicasting. In the future, I am supposed to work on the continuation of the paper titled "Bootstrapping The Long Tail of Peer-to-Peer Systems" by B. Huberman and F. Wu. I will try to summarize it briefly and ask some questions about how I should proceed.

1. Leslie Fine

  • Title:
  • Abstract:
DateTimeLocation
7/2010AM-12PM Kappa

1. Saikat Guha

  • Previous Work: Capacities of quantum optical communication channels, quantum convolutional error-correcting codes (http://web.mit.edu/saikat/www/research/)
  • Summer Plan: Studying cheating mechanisms and implementation issues of quantum auctions
  • Abstract: A protocol to run auctions that uses quantum superpositions to represent bids and quantum adiabatic search to identify the winning bid was proposed, where a final projective measurement on the overall quantum state identifies the winning bids and simultaneously destroys the bidding states, thereby preserving privacy. Correlated bids can be realized using quantum entanglement. [Hogg, Chen, Beausoleil, Spiller] (1) We would like to investigate ways in the auctions protocol, by which either dishonest bidders or a corrupt auctioneer can cheat in order to motivate their respective incentives, which can for instance be, for bidders -- getting less than highest bids to win the auctions with high probability, and for an auctioneer – to collect bid values and bidding patterns of bidders. We would like to find out ways to improve the protocol by which cheating can be either caught or prevented. (2) We are also interested in implementing auctions using optical quantum computing, as this protocol can serve as a simple non-trivial demonstration of quantum computing using a few qubits. We would like to investigate into implementations that only make use of two-qubit unitary operations.

1. Dennis Wilkinson

  • Title: TBA
  • Abstract: TBA

1. Tad Hogg

  • Title: Taking risk away from risk taking: Decision insurance in organizations
  • Abstract: I'll first give a brief overview of my various interests in using new technologies, such as availablity of social networks, for designing economic mechanisms.

Then I'll focus on one example: decision insurance, an insurance-like mechanism for encouraging risk taking by decision-makers within organizations. Since insurance increases the likelihood of free riding, we also introduce a technique that mitigates this moral hazard by automatically identifying the social network around the decision-makers and using it as a monitoring group. We show that three possible regimes exist. In the first one, people contribute to production but avoid risky projects. In the second, people take on risky projects without free riding. In the third, they free ride. We establish the conditions for the appearance of each of these regimes and show how to adjust the mechanism parameters so as to get the highest expected payoff for the firm in spite of its risk-adverse members. We plan laboratory tests of this mechanism later in 2006, which will also give us experience for designing experiments for other mechanisms involving social networks, such as reputation.

DateTimeLocation
7/2710am-12pmTioga
DateTimeLocation
8/3No Meeting n/a
8/10No Meeting n/a
8/1710-12AM Kappa

1. Kevin Lai

  • Title: Using Contingency Funds to Mitigate Risk
  • Abstract: Market-based Proportional Share scheduling provides agile and fair scheduling of shared resources while providing little predictability. Auctions of reserved slots allows predictable scheduling at the cost of susceptibility to gaming and high latency. This talk is about contingency scheduling, which combines the agility of proportional share scheduling with controllable predictability. I'll describe the basic algorithm and tradeoffs of contingency scheduling, including signalling and fraud.

1. Li Zhang

  • Title: TBA
  • Abstract: TBA
DateTimeLocation
8/2410-12AM Tioga

1. Dave Levin

  • Title: Cold-Start Multicast
  • Abstract: We explore the problem of multicasting to a potentially large number of recipients who have not opted in to receive. To be of use in, say, a PlanetLab?-like environment, such a tool requires certain features, especially scalability, speed, feedback, and authentication. I will discuss our solutions to these, their inherent trade-offs, and present some preliminary results from this summer. I will also discuss proposed future work.

1. Thomas Sandholm

  • Title: Prediction-Based QoS Guarantees in a Proportional Share Market
  • Abstract: In this talk I will present work on using price predictions to offer resource performance-level guarantees. The trade-off between price and guarantee will be discussed in the context of different work-load scenarios and simulation results will be shown. Statistical guarantees will be compared to a new proposed technique which implements admission control on top of predictions to increase the guarantee-level when there is resource contention.

8/3110-12Kappa

1. Josep Pujol

  • Title:
  • Abstract:
9/710-12AM Tioga

1. Alvin AuYoung

  • Title:
  • Abstract:
9/1410-12AMKappa

1. Ali Ghodsi

  • Title: The Survival of the Unfittest
  • Abstract: Previous work suggested applying pari-mutuel markets to peer-to-peer content distribution networks. This talk focuses on the shape of the equilibriums that arise when item popularities are ZIPF distributed. We investigate the effects of bounded rationality and limited view.
9/2110-12AMTioga

1. Travis Kriplean

  • Title: Tracking Norms in Wikipedia via Policies
  • Abstract: I describe my summer work examining the maturation of the wikipedian community through the guise of policy and keyword citations on talk pages. The study highlights the increasing prominence of policies throughout Wikipedia's history as well as the growing influence of a small number of users with "adminstrator" status. Unfortunately, due to the lack of qualitative investigation to back up the quantitative metrics, it is difficult make grounded statements about the power relations that come into play when the policies and keywords are being used--this is future work.