Aditya Kurve

Optimizing cluster formation in super-peer networks via local incentive design

A super-peer based overlay network architecture for peer-to-peer (P2P) systems allows for some nodes, known as the super-peers, that are more resource-endowed than others, to assume a higher share of workload. Ordinary peers are connected to the super-peers and rely on them for their transactional needs. Many criteria for a peer to choose its super-peer have been explored, some of them based on physical proximity, semantic proximity, or by purely random choice. In this paper, we propose an incentive-based criterion that uses semantic similarities between the content interests of the peers and, at the same time, encourages even load distribution across the super-peers. The incentive is achieved via a game theoretic framework that considers each peer as a rational player, allowing stable Nash equilibria to exist and hence guarantees a fixed point in the strategy space of the peers. This guarantees convergence (assuming static network parameters) to a locally optimal assignment of peers to super-peers with respect to a global cost that approximates the average query resolution time. We also show empirically that the local cost framework that we employ performs closely to (and in some cases better than) a similar scheme based on the formulation of a centralized cost function that requires the peers to know an additional global parameter.

Multicategory Crowdsourcing Accounting for Variable Task Difficulty, Worker Skill, and Worker Intention

Crowdsourcing allows instant recruitment of workers on the web to annotate image, webpage, or document databases. However, worker unreliability prevents taking a workers responses at “face value”. Thus, responses from multiple workers are typically aggregated to more reliably infer ground-truth answers. We study two approaches for crowd aggregation on multicategory answer spaces: stochastic modeling-based and deterministic objective function-based. Our stochastic model for answer generation plausibly captures the interplay between worker skills, intentions, and task difficulties and captures a broad range of worker types. Our deterministic objective-based approach aims to maximize the average aggregate confidence of weighted plurality crowd decision making. In both approaches, we explicitly model the skill and intention of individual workers, which is exploited for improved crowd aggregation. Our methods are applicable in both unsupervised and semi-supervised settings, and also when the batch of tasks is heterogeneous, i.e., from multiple domains, with task-dependent answer spaces. As observed experimentally, the proposed methods can defeat “tyranny of the masses”, i.e., they are especially advantageous when there is an (a priori unknown) minority of skilled workers amongst a large crowd of unskilled (and malicious) workers.

An agent-based framework for performance modeling of an optimistic parallel discrete event simulator

PurposeThe performance of an optimistic parallel discrete event simulator (PDES) in terms of the total simulation execution time of an experiment depends on a large set of variables. Many of them have a complex and generally unknown relationship with the simulation execution time. In this paper, we describe an agent-based performance model of a PDES kernel that is typically used to simulate large-sized complex networks on multiple processors or machines. The agent-based paradigm greatly simplifies the modeling of system dynamics by representing a component logical process (LP) as an autonomous agent that interacts with other LPs through event queues and also interacts with its environment which comprises the processor it resides on.MethodWe model the agents representing the LPs using a “base” class of an LP agent that allows us to use a generic behavioral model of an agent that can be extended further to model more details of LP behavior. The base class focuses only on the details that most likely influence the overall simulation execution time of the experiment.ResultsWe apply this framework to study a local incentive based partitioning algorithm where each LP makes an informed local decision about its assignment to a processor, resulting in a system akin to a self organizing network. The agent-based model allows us to study the overall effect of the local incentive-based cost function on the simulation execution time of the experiment which we consider to be the global performance metric.ConclusionThis work demonstrates the utility of agent-based approach in modeling a PDES kernel in order to evaluate the effects of a large number of variable factors such as the LP graph properties, load balancing criteria and others on the total simulation execution time of an experiment.

Sybil Detection via Distributed Sparse Cut Monitoring

Decentralized reputation systems help to enforce discipline and fairness in large unstructured and ad-hoc systems by rewarding good behavior and penalizing dishonest or greedy behavior. They are essential in large networks of independent nodes where centralized monitoring of node behavior is difficult due to the sheer size of the network. Sybil nodes pose a threat to the reputation systems by false referrals through sybil identities. We propose a scalable and distributed algorithm to identify attack edges and quarantine sybil clusters. This algorithm works well with dynamic trust graphs as nodes do not need to store any pre-computed data.

Iterative partitioning scheme for distributed simulation of dynamic networks

Network model partitioning is a key component of distributed network simulations. Simulations slow down considerably due to inequitable load balancing and heavy inter-host communication leading to unbounded synchronization overhead. Also, regularly refreshing the node partition is necessary due to to the dynamic nature of simulation load and event generation. In this paper, we propose a distributed method for network partitioning which includes a coarse initial partitioning followed by iterative improvements in the partition. We suggest a sparse-cut based method to identify nodes eligible for exchange.

A study of unsupervised adaptive crowdsourcing

We consider unsupervised crowdsourcing performance based on the model wherein the responses of end-users are essentially rated according to how their responses correlate with the majority of other responses to the same subtasks/questions. In one setting, we consider an independent sequence of identically distributed crowdsourcing assignments (meta-tasks), while in the other we consider a single assignment with a large number of component subtasks. Both problems yield intuitive results in which the overall reliability of the crowd is a factor.

Decentralized capacity reallocation for a loss network

Stochastic loss network models have been widely used to study the characteristics of a large class of resource constrained networks in terms of their blocking probability. We consider a distributed capacity allocation problem for a loss network with directed edges. Each infrastructure node (or superpeer) can independently adjust its link-capacity allocations, subject to a constraint on the total amount for each node, so as to minimize its estimate of call blocking rates. We argue via Erlang fixed-point approximation that such decentralized local changes do work to minimize a global measure of weighted call blocking rates.