30.01.2014 Views

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Aspects of Wardrop Equilibria<br />

Lars Olbrich<br />

Global communication networks like the Internet often lack a central authority that monitors<br />

<strong>an</strong>d regulates network traffic. Network users may behave selfishly according to their private<br />

interest without regard to the overall system perform<strong>an</strong>ce. Such highly complex environments<br />

prompted a paradigm shift in computer science. Whereas traditional concepts are designed for<br />

st<strong>an</strong>d-alone machines <strong>an</strong>d m<strong>an</strong>ageable networks, a profound un<strong>der</strong>st<strong>an</strong>ding of large-scale<br />

communication systems with strategic users requires to combine methods from theoretical<br />

computer science with game-theoretic techniques. In this thesis, we study equilibrium<br />

situations in Wardrop's traffic model. In Wardrop's model a rate of traffic between each pair<br />

of vertices of a network is modeled as network flow, i.e., traffic is allowed to split into<br />

arbitrary pieces. The resources are the network edges with latency functions qu<strong>an</strong>tifying the<br />

time needed to traverse <strong>an</strong> edge. The latency of <strong>an</strong> edge depends on the congestion. It<br />

increases the more flow traverses that edge. A common interpretation of the Wardrop model<br />

is that flow is controlled by <strong>an</strong> infinite number of agents each of which is responsible to route<br />

<strong>an</strong> infinitesimal amount of traffic between its origin <strong>an</strong>d destination vertex. Each agent plays a<br />

pure strategy in choosing one path from its origin to its destination, where the agent's<br />

disutility is the sum of edge latencies on this path. A Wardrop equilibrium denotes a strategy<br />

profile in which all used paths between a given origin-destination pair have equal <strong>an</strong>d<br />

minimal latency. Wardrop equilibria are also Nash equilibria as no agent c<strong>an</strong> decrease its<br />

experienced latency by unilaterally deviating to <strong>an</strong>other path. Like Nash equilibria in general,<br />

Wardrop equilibria do not optimize <strong>an</strong>y global objective per se. In particular, the total latency<br />

of all agents is not minimized at Wardrop equilibrium. Addressing this issue, Roughgarden<br />

<strong>an</strong>d Tardos gave tight bounds on the price of <strong>an</strong>archy measuring the worst-possible<br />

inefficiency of equilibria with respect to the incurred latency. Further, the famous Braess's<br />

paradox states that adding edges to a network may in fact worsen the unique equilibrium. The<br />

primary goal of this thesis is to provide a deeper un<strong>der</strong>st<strong>an</strong>ding of Wardrop equilibria. We<br />

identify several problems whose solution captures the essence of Wardrop equilibria. First,<br />

we study natural <strong>an</strong>d innovative me<strong>an</strong>s to reduce the price of <strong>an</strong>archy. Secondly, we <strong>an</strong>alyze<br />

the stability of equilibria regarding modifications of the network environment. Finally, we<br />

propose a distributed algorithm for computing approximate equilibria. The inefficiency of<br />

equilibria motivates our first line of research. In Wardrop's model, imposing marginal cost<br />

taxes on every edge completely eliminates the inefficiency of selfish routing. We concentrate<br />

on optimal taxes for the crucial <strong>an</strong>d more realistic case in which only a given subset of the<br />

edges c<strong>an</strong> be taxed. We establish NP-hardness of this optimization problem in general<br />

networks. On the positive side, we provide a polynomial time algorithm for computing<br />

optimal taxes in parallel link networks with linear latency functions. We also propose a novel<br />

approach to improve the perform<strong>an</strong>ce of selfish flow in networks by additionally routing<br />

flow, called auxiliary flow. We focus on the computational complexity for the optimal<br />

utilization of auxiliary flow <strong>an</strong>d present strong inapproximability results. In particular, the<br />

minimal amount of auxiliary flow needed to induce <strong>an</strong> optimal flow as the outcome of selfish<br />

behavior c<strong>an</strong>not be approximated by <strong>an</strong>y subexponential factor. Further, we study the<br />

sensitivity of Wardrop equilibria. From both the practical <strong>an</strong>d the theoretical perspective it is<br />

a natural <strong>an</strong>d intriguing question, how equilibria respond to slight modifications of either the<br />

network topology or the traffic flow. We show positive <strong>an</strong>d negative results on the stability of<br />

flow pattern <strong>an</strong>d flow characteristics at equilibrium. As it is fundamental for the above studies<br />

that selfish behavior in network routing games yields <strong>an</strong> equilibrium, it is not clear how the<br />

set of agents c<strong>an</strong> attain <strong>an</strong> equilibrium in the first place. In previous work it was shown that <strong>an</strong><br />

485

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!