DSS Schedule for Spring 2012

With the advent of industrial standards such as WirelessHART, process industries are now gravitating towards wireless control systems. Due to limited bandwidth in a wireless network shared by multiple control loops, it is critical to optimize the overall control performance. In this paper, we address the scheduling-control co-design problem of determining the optimal sampling rates of feedback control loops sharing a WirelessHART network. The objective is to minimize the overall control cost while ensuring that all data flows meet their end-to-end deadlines. The resulting constrained optimization based on existing delay bounds for WirelessHART networks is challenging since it is non-differentiable, non-linear, and not in closed-form. We propose four methods to solve this problem. First, we present a subgradient method for rate selection. Second, we propose a greedy heuristic that usually achieves low control cost while significantly reducing the execution time. Third, we propose a global constrained optimization algorithm using a simulated annealing (SA) based penalty method. Finally, we formulate rate selection as a differentiable convex optimization problem that provides a closed-form solution through a gradient descent method. This is based on a new delay bound that is convex and differentiable, and hence simplifies the optimization problem. We evaluate all methods through simulations based on topologies of a 74-node wireless sensor network testbed. Surprisingly, the subgradient method is disposed to incur the longest execution time as well as the highest control cost among all methods. SA and the greedy heuristic represent the opposite ends of the tradeoff between control cost and execution time, while the gradient descent method hits the balance between the two.

Content distribution is a primary activity on the Internet. Name-centric network architectures support content distribution intrinsically. Named Data Networking (NDN), one recent such proposal, names packets rather than end-hosts, thereby enabling packets to be cached and redistributed by routers. Among alternative name-based systems, HTTP is the most significant by any measure. A majority of today's content distribution services leverage widely deployed HTTP infrastructure, such as web servers and caching proxies. As a result, HTTP can be viewed as a practical, name-based content distribution solution. Of course, NDN and HTTP do not overlap entirely in their capabilities and design goals, but they do overlap in the area of name-based content distribution. This talk presents an experimental performance evaluation of NDN-based and HTTP-based content distribution solutions. Specifically, the NDN-based method employs CCNx, an NDN reference software implementation, while the HTTP-based solution leverages popular web server lighttpd and the Squid caching web proxy. All three systems are available as open-source software. Using a reconfigurable network testbed, we are able to base our measurements on running code on real hardware systems, over a range of controlled and repeatable experimental conditions. Our findings verify popular intuition, but also surprise in some ways. In wired networks with local-area transmission latencies, the HTTP-based solution dramatically outperforms NDN, with roughly 10x greater sustained throughput. In networks with lossy access links, such as wireless links with 5% drop rates, or with non-local transmission delays, the situation reverses and NDN outperforms HTTP, with sustained throughput by increased by roughly 4x over a range of experimental scenarios.

Large Margin Nearest Neighbors (LMNN) is an effective algorithm for metric learning, especially designed to reduces the k-nearest neighbors classification error. Similar to most recently published metric learning algorithms, LMNN learns a Mahalanobis metric. Although the Mahalanobis metric allows the training to be formulated as an elegant semi-definite program, it also comes with the inherent limitation of being merely an affine transformation of the input space.  In this talk I present an extension to LMNN, where the linear mapping is substituted by an ensemble of non-linear gradient boosted trees. Different from previous approaches to ``go beyond'' linear mappings, which almost entirely focussed on kernelization, this approach scales linearly with the size of the training data. Experimental results validate that our non-linear extension to LMNN leads to better k-nearest neighbor classification results especially when combined with dimensionality reduction. 

The intensity-modulated radiation therapy treatment planning (IMRT) is a computationally ex- pensive problem because of a large number of constraints and variables in it. These problems are computationally expensive requiring a lot of time and memory space. In this research, we propose a new sampling method called Computational Boundary Sampling, which reduces the time and space required to solve the IMRT problem. In this sampling method, we distinguished between the boundary and inner-region of an organ. The problem formulation includes all the voxels from the boundary-region, but includes only 'p' percent from the inner-region. We found that this method significantly reduces the time and memory required to solve the problem without degrading the final dose distribution of organs.

To facilitate long-term engagement in programming for middle school children, we developed the Looking Glass community. The community includes both a website and integrated access to community resources within the Looking Glass programming environment. It supports engagement by providing a source for initial ideas, support for learning new skills, positive feedback, and role models.

Wireless sensor networks are evolving from single- application platforms towards an integrated infrastructure shared by multiple applications. Given the resource constraints of sensor nodes, it is important to optimize the allocation of applications to maximize the overall Quality of Monitoring (QoM). Recent solutions to this challenging application allocation problem are centralized in nature, limiting their scalability and robustness against network failures and dynamics. This paper presents a distributed game-theoretic approach to application allocation in shared sensor networks. We first transform the optimal application allocation problem to a submodular game and then develop a decentralized algorithm that only employs localized interactions among neighboring nodes. We prove that the network can converge to a pure strategy Nash equilibrium with an approximation bound of 1/2. Simulations based on three real-world datasets demonstrate that our algorithm is competitive against a state-of-the-art centralized algorithm in terms of QoM.

ScalaPipe is a streaming application generator for heterogeneous platforms. By using a collection of domain-specific languages (DSLs) embedded in the Scala programming language, ScalaPipe allows creation of streaming applications that can run on a variety of hardware, including traditional processors, graphics processors, and field-programmable gate arrays (FPGAs). Its application DSL allows specification of the application topology and resource mapping. Its block DSL allows the authoring of implementations for processing kernels, or blocks, which are used in the streaming application. ScalaPipe makes it easy to generate, modify, and instrument large, complex topologies and resource mappings while also exposing optimization opportunities.

Transmission power control (TPC) of wireless Sensor Network (WSN)has gained considerable attention in recent years. It faces significant challenges like fluctuation of link quality especially in indoor environment, resources constrained hardware platforms. Although there are a number of Previous researches to tackle the problem of transmission control for indoor environment. They suffer drawbacks such as inconsistent link quality index, computation intensive algorithms. To meet the challenges and overcome the shortcomings of previous work, we propose a novel algorithm of Feedback Efficient Transmission Power Control (FETPC). FETPC features a simple and efficient controller that integrates a proportional-integral (PI) and a traditional anti-windup (AW) controller designed to enforce the hardware limits of transmission power. To accommodate the significant disturbance, a fast online model identification algorithm is also developed and added to the feedback loop to facilities the reconfiguration of the controller. Extensive simulation results demonstrate FETPC can achieve better performance than the algorithms with simple heuristic and those algorithms based on inconsistent performance index. Meanwhile it also has lower overhead than that of computation intensive algorithms.

Low Power Listening (LPL) is a common MAC-layer technique for reducing energy consumption in wireless sensor networks, where nodes periodically wake up to sample the wireless channel to detect activity. However, LPL is highly susceptible to false wakeups caused by environmental noise being detected as activity on the channel, causing nodes to spuriously wake up in order to receive nonexistent transmissions. In empirical studies in residential environments, we observe that the false wakeup problem can significantly increase a node's duty cycle, compromising the benefit of LPL. We also find that the energy-level threshold used by the Clear Channel Assessment (CCA) mechanism to detect channel activity has a significant impact on the false wakeup rate. We then design AEDP, an adaptive energy detection protocol for LPL, which dynamically adjust a node's CCA threshold to meet application-specified bounds on network reliability and duty cycle. Empirical experiments in both controlled tests and residential environments showed AEDP can effectively mitigate the impact of noise on radio duty cycles, while maintaining satisfactory link reliability.

A reactive program is one that has “ongoing interactions with its environment”[1]. Reactive programs are the foundation of all embedded, networked, and interactive applications. Reactive programs are difficult to develop due to inherent concurrency and asynchrony. Furthermore, existing platforms are not designed to support reactive programs but instead have been retrofitted with reactive semantics, leading to unnecessary accidental complexity. In this work, we take the position that a system that allows developers to implement reactive programs directly is necessary to facilitate the development of reactive systems of increasing complexity. To this end, we present Lily, a microkernel for reactive programs based on the I/O automata formal model. Developers using Lily write reactive components that are then bound together at run-time to form dynamic reactive systems. We demonstrate these features of Lily with a simple software system that samples a free-running timer and a test program for a PS/2 mouse driver. 1. Manna, Z. and Pnueli, A., “The temporal logic of reactive and concurrent systems: Specification.”

Neural networks have been a fundamental model in artificial intelligence and machine learning for decades. Recent successes of "deep learning" techniques established neural networks among the most powerful learning methods. Increases in computing speed made training large networks feasible, and better structural intuitions set new benchmarks for prediction accuracy. However, further improvements in training time are needed and must be accomplished by harnessing parallel systems. The typical training method, error back-propagation, is highly sequential. In this work, we demonstrate a relaxed back-propagation algorithm permitting a pipeline training method that scales with neural network depth. The pipeline method maintains high accuracy while offering significant speedups for deep neural networks.

Presentation material has been removed because it will be submitted for publication.

Wireless sensor networks will play an important role improving patient care by collecting continuous vital signs for clinical decision support. This paper presents the architecture and experience with a large-scale wireless clinical monitoring system designed to support detecting clinical deterioration among hospitalized patients. Our system integrates portable wireless pulse oximeters attached to patients, a wireless relay network spanning multiple hospital floors and wards and an Electronic Medical Records (EMR). We report our experience and lessons learned from the deployment and maintenance of our system in four hospital wards on three floors of Barnes-Jewish Hospital over a year long clinical trial. Our experience showed the feasibility to achieve reliable vital sign collection using a wireless sensor network integrated with the hospital IT infrastructure, and highlight technical and non-technical aspects that pose challenges in a complex hospital environment. We conclude with a list of DOs and DON’Ts that will enable successful and efficient deployment of wireless sensor networks in clinical environments.