Document Sample
outline Powered By Docstoc

A couple of challenging research problems
Dec 1, 2007 Tsinghua-CUHK Workshop

Dah Ming Chiu Chinese University of Hong Kong

Beyond “network utility maximization” Modeling the academic publishing game M



Network resource allocation
Case 1: single broadcast link
N nodes have packets to send, how to share the link? d h k d h h h li k? A scheduling problem ALOHA algorithms, each node sends with probability p Not that optimal, but distributed, robust, simple, fair The best p is a function of N – find best p adaptively: exponential b k ff algorithms i l back-off l i h Fairness is self-evident Follow-up work: 1000s of papers

Network resource allocation
Case 2: flows sharing multiple link network
N flows (source-dest) have packets to send, how to allocation bandwidth of each link? Solution: “AIMD” – TCP congestion control Not that perfect in many situations, but distributed, robust, simple, fair Fairness not that obvious





Frank Kelly’s utility maximization model
Each user has a convex utility function (of flow rate) ) Maximize total utility of all flows, subject to link bandwidth constraints Established meaning of fairness Showed AIMD/TCPCC is but a distributed algorithm to solve the utility maximization problem Studied stability Established linkage to congestion price Follow-up work: 1000s of papers




Limitations to utility maximization
It is a qualitative framework to express policy (fairness) In real networks, some problems:
Different utility functions for different applications e.g. inelastic flows may use other traffic controls Users are not charged for each flow – so to get more resources, users can use more flows fl e.g. p2p downloading -> fairness is meaningless



A new theory of network resource allocation?
We are considering different models towards this end Utility maximization f users with different utility for ff wrote a paper on traffic controls for inelastic flows ISP peering based on Shapley Values – collaboration with Columbia Univ, and Jian Wei Credit-based networking – collaboration with Prof Wang Ji Long of Tsinghua University Performance evaluation of Paris Metro Pricing – current graduate student project

Modeling the academic publishing game
As a community, we are publishing more and more papers
The selection process is becoming more random We cannot catch up with all the papers published, even those related to our work; there are many duplications Many games people play, to advance their publishing career

How can we improve the situation?



Better reviewing process
Performance modeling of conference review process
Similar Si il to evaluation of P2P systems? l i f ? Any classic algorithm for ranking items applicable? How about a community-based public ranking system – all papers “published” at a pre-print server and subjected to public review?
Collusion problem Lack of incentive to review others’ work The problem of incremental publication

Perfect application for game theory
Does the current equilibrium give too much emphasis to q antit rather than q alit ? quantity quality? Does the current system encourage too many names on the same paper?



Data mining of academic publishing
Mining paper databases
Mining Mi i EDAS (Collaboration with P f S h l i (C ll b i i h Prof Schulzrinne of f Columbia University) to study trends of conferences, conferences papers and collaboration patterns First phase – FYP mining various conferences available on-line

Mining Google Scholar
Past FYP – how to deal with name collision problem, and others


Shared By:
Tags: outli
Description: outline