Christos G. Panayiotou

Perturbation analysis for online control and optimization of stochastic fluid models

Uses stochastic fluid models (SFMs) for control and optimization (rather than performance analysis) of communication networks, focusing on problems of buffer control. We derive gradient estimators for packet loss and workload related performance metrics with respect to threshold parameters. These estimators are shown to be unbiased and directly observable from a sample path without any knowledge of underlying stochastic characteristics, including traffic and processing rates (i.e., they are nonparametric). This renders them computable in online environments and easily implementable for network management and control. We further demonstrate their use in buffer control problems where our SFM-based estimators are evaluated based on data from an actual system.

Perturbation analysis and control of two-class stochastic fluid models for communication networks

This paper uses stochastic fluid models (SFMs) for the control and optimization (rather than performance analysis) of communication network nodes processing two classes of traffic: one is uncontrolled and the other is subject to threshold-based buffer control. We derive gradient estimators for packet loss and workload related performance metrics with respect to threshold parameters. These estimators are shown to be unbiased and directly observable from a sample path without any knowledge of underlying stochastic characteristics of the traffic processes. This renders them computable in online environments and easily implementable for network management and control. We further demonstrate their use in buffer control problems where our SFM-based estimators are evaluated based on data from an actual system.

Perturbation Analysis and Optimization of Stochastic Hybrid Systems

We present a general framework for carrying out perturbation analysis in Stochastic Hybrid Systems (SHS) of arbitrary structure. In particular, Infinitesimal Perturbation Analysis (IPA) is used to provide unbiased gradient estimates of performance metrics with respect to various controllable parameters. These can be combined with standard gradient-based algorithms for optimization purposes and implemented on line with little or no distributional information regarding the stochastic processes involved. We generalize an earlier concept of “induced events” for this framework to include system features such as delays in control signals or modeling multiple user classes sharing a resource. We apply this generalized IPA to two SHS with different characteristics. First, we develop a gradient estimator for the performance of a linear switched system with control signal delays and a safety constraint and show that it is independent of the random delays distributional characteristics. Second, we derive closed-form unbiased IPA estimators for a Stochastic Flow Model (SFM) of systems executing tasks subject to either hard or soft real-time constraints. These estimators are incorporated in a gradient-based algorithm to optimize performance by controlling a task admission threshold parameter. Simulation results are included to illustrate this optimization approach.

Online IPA Gradient Estimators in Stochastic Continuous Fluid Models

This paper applies infinitesimal perturbation analysis (IPA) to loss-related and workload-related metrics in a class of stochastic flow models (SFM). It derives closed-form formulas for the gradient estimators of these metrics with respect to various parameters of interest, such as buffer size, service rate, and inflow rate. The IPA estimators derived are simple and fast to compute, and are further shown to be unbiased and nonparametric, in the sense that they can be computed directly from the observed data without any knowledge of the underlying probability law. These properties hold out the promise of utilizing IPA gradient estimates as ingredients of online management and control of telecommunications networks. While this paper considers single-node SFMs, the analysis method developed is amenable to extensions to networks of SFM nodes with more general topologies.

Machine learning competition in immunology – Prediction of HLA class I binding peptides

Experimental studies of immune system and related applications such as characterization of immune responses against pathogens, vaccine design, or optimization of therapies are combinatorially complex, time-consuming and expensive. The main methods for large-scale identification of T-cell epitopes from pathogens or cancer proteomes involve either reverse immunology or high-throughput mass spectrometry (HTMS). Reverse immunology approaches involve pre-screening of proteomes by computational algorithms, followed by experimental validation of selected targets (Mora et al., 2006; De Groot et al., 2008; Larsen et al., 2010). HTMS involves HLA typing, immunoaffinity chromatography of HLA molecules, HLA extraction, and chromatography combined with tandem mass spectrometry, followed by the application of computational algorithms for peptide characterization (Bassani-Sternberg et al., 2010). Hundreds of naturally processed HLA class I associated peptides have been identified in individual studies using HTMS in normal (Escobar et al., 2008), cancer (Antwi et al., 2009; Bassani-Sternberg et al., 2010), autoimmunity-related (Ben Dror et al., 2010), and infected samples (Wahl et al, 2010). Computational algorithms are essential steps in highthroughput identification of T-cell epitope candidates using both reverse immunology and HTMS approaches. Peptide binding to MHC molecules is the single most selective step in defining T cell epitope and the accuracy of computational algorithms for prediction of peptide binding, therefore, determines the accuracy of the overall method. Computational predictions of peptide binding to HLA, both class I and class II, use a variety of algorithms ranging from binding motifs to advanced machine learning techniques (Brusic et al., 2004; Lafuente and Reche, 2009) and standards for their

Adaptive Approximation for Multiple Sensor Fault Detection and Isolation of Nonlinear Uncertain Systems

This paper presents an adaptive approximation-based design methodology and analytical results for distributed detection and isolation of multiple sensor faults in a class of nonlinear uncertain systems. During the initial stage of the nonlinear system operation, adaptive approximation is used for online learning of the modeling uncertainty. Then, local sensor fault detection and isolation (SFDI) modules are designed using a dedicated nonlinear observer scheme. The multiple sensor fault isolation process is enhanced by deriving a combinatorial decision logic that integrates information from local SFDI modules. The performance of the proposed diagnostic scheme is analyzed in terms of conditions for ensuring fault detectability and isolability. A simulation example of a single-link robotic arm is used to illustrate the application of the adaptive approximation-based SFDI methodology and its effectiveness in detecting and isolating multiple sensor faults.

Perturbation analysis and optimization of stochastic flow networks

We consider a stochastic fluid model of a network consisting of several single-class nodes in tandem and perform perturbation analysis for the node queue contents and associated event times with respect to a threshold parameter at the first node. We then derive infinitesimal perturbation analysis (IPA) derivative estimators for loss and buffer occupancy performance metrics with respect to this parameter and show that these estimators are unbiased. We also show that the estimators depend only on data directly observable from a sample path of the actual underlying discrete event system, without any knowledge of the stochastic characteristics of the random processes involved. This renders them computable in online environments and easily implementable for network management and optimization. This is illustrated by combining the IPA estimators with standard gradient based stochastic optimization methods and providing simulation examples.

A Low-Cost Sensor Network for Real-Time Monitoring and Contamination Detection in Drinking Water Distribution Systems

This paper presents a low cost and holistic approach to the water quality monitoring problem for drinking water distribution systems as well as for consumer sites. Our approach is based on the development of low cost sensor nodes for real time and in-pipe monitoring and assessment of water quality on the fly. The main sensor node consists of several in-pipe electrochemical and optical sensors and emphasis is given on low cost, lightweight implementation, and reliable long time operation. Such implementation is suitable for large scale deployments enabling a sensor network approach for providing spatiotemporally rich data to water consumers, water companies, and authorities. Extensive literature and market research are performed to identify low cost sensors that can reliably monitor several parameters, which can be used to infer the water quality. Based on selected parameters, a sensor array is developed along with several microsystems for analog signal conditioning, processing, logging, and remote presentation of data. Finally, algorithms for fusing online multisensor measurements at local level are developed to assess the water contamination risk. Experiments are performed to evaluate and validate these algorithms on intentional contamination events of various concentrations of escherichia coli bacteria and heavy metals (arsenic). Experimental results indicate that this inexpensive system is capable of detecting these high impact contaminants at fairly low concentrations. The results demonstrate that this system satisfies the online, in-pipe, low deployment-operation cost, and good detection accuracy criteria of an ideal early warning system.

SNAP: Fault Tolerant Event Location Estimation in Sensor Networks Using Binary Data

This paper investigates the use of wireless sensor networks for estimating the location of an event that emits a signal that propagates over a large region. In this context, we assume that the sensors make binary observations and report the event (positive observations) if the measured signal at their location is above a threshold; otherwise, they remain silent (negative observations). Based on the sensor binary beliefs, a likelihood matrix is constructed whose maximum value points to the event location. The main contribution of this work is Subtract on Negative Add on Positive (SNAP), an estimation algorithm that provides an efficient way of constructing the likelihood matrix by simply adding \pm 1 contributions from the sensor nodes depending on their alarm state (positive or negative). This simple estimation procedure provides very accurate results and turns out to be fault tolerant even when a large percentage of the sensor nodes report erroneous observations.

Crowdsourced indoor localization for diverse devices through radiomap fusion

Crowdsourcing is an emerging field that allows to tackle difficult problems by soliciting contributions from common people, rather than trained professionals. In the post-pc era, where smartphones dominate the personal computing market offering both constant mobility and large amounts of spatiotemporal sensory data, crowdsourcing is becoming increasingly popular. In this context, crowdsourcing stands as the only viable solution for collecting the large amount of location-related network data required to support location-based services, e.g., the signal strength radiomap of a fingerprinting localization system inside a multi-floor building. However, this benefit does not come for free, because crowdsourcing also poses new challenges in radiomap creation. We focus on the problem of device diversity that occurs frequently as the contributors usually carry heterogeneous mobile devices that report network measurements very differently. We demonstrate with simulations and experimental results that the traditional signal strength values from the surrounding network infrastructure are not suitable for crowdsourcing the radiomap. Moreover, we present an alternative approach, based on signal strength differences, that is far more robust to device variations and maintains the localization accuracy regardless of the number of contributing devices.