Doina Bein

Distributed MPI cluster with Docker Swarm mode

MPI is a well-established technology that is used widely in high-performance computing environment. However, setting up an MPI cluster can be challenging and time-consuming. This paper tackles this challenge by using modern containerization technology, which is Docker, and container orchestration technology, which is Docker Swarm mode, to automate the MPI cluster setup and deployment. We created a ready-to-use solution for developing and deploying MPI programs in a cluster of Docker containers running on multiple machines, orchestrated with Docker Swarm mode, to perform high computation tasks. We explain the considerations when creating Docker image that will be instantiated as MPI nodes, and we describe the steps needed to set up a fully connected MPI cluster as Docker containers running in a Docker Swarm mode. Our goal is to give the rationale behind our solution so that others can adapt to different system requirements. All pre-built Docker images, source code, documentation, and screencasts are publicly available.

Algorithmic Approaches for a Dependable Smart Grid

We explore options for integrating sustainable and renewable energy into the existing power grid, or even create a new power grid model. We present various theoretical concepts necessary to meet the challenges of a smart grid. We first present a supply and demand model of the smart grid to compute the average number of conventional power generator required to meet demand during the high consumption hours. The model will be developed using Fluid Stochastic Petri Net (FSPN) approach. We propose to model the situations that need decisions to throttle down the energy supplied by the traditional power plants using game-theoretic online competitive models. We also present in this paper the power-down model which has shown to be competitive in the worst case scenarios and we lay down the ground work for addressing the multi-state dynamic power management problem.

Low-cost, real-time obstacle avoidance for mobile robots

The goal of this project1 is to advance the field of automation and robotics by utilizing recently-released, low-cost sensors and microprocessors to develop a mechanism that provides depth-perception and autonomous obstacle avoidance in a plug-and-play fashion. We describe the essential hardware components that can enable such a low-cost solution and an algorithm to avoid static obstacles present in the environment. The mechanism utilizes a novel single-point LIDAR module that affords more robustness and invariance than popular approaches, such as Neural Networks and Stereo. When this hardware is coupled with the proposed efficient obstacle avoidance algorithm, this mechanism is able to accurately represent environments through point clouds and construct obstacle-free paths to a destination, in a small timeframe. A prototype mechanism has been installed on a quadcopter for visualization on how actual implementation may take place2. We describe experimental results based on this prototype.

Securing unmanned autonomous systems from cyber threats

Unmanned systems, with and without a human-in-the loop, are being deployed in a range of military and civilian applications spanning air, ground, sea-surface and undersea environments. Large investments, particularly in robotics, electronic miniaturization, sensors, network communication, information technology and artificial intelligence are likely to further accelerate this trend. The operation of unmanned systems, and of applications that use these systems, are heavily dependent on cyber systems that are used to collect, store, process and communicate data, making data a critical resource. At the same time, undesirable elements of our society and adversarial states have also realized the high value of this resource. While enormous efforts have been made to secure data and cyber systems, lack of rigorous threat modeling and risk analysis can lead to more specific, rather than generic, security solutions relevant to the cyber system to be protected. This scenario has created an urgent need to develop a holistic process for protecting data and cyber systems. This paper deals with the development of different pieces of this process. We first identify the security requirements of unmanned autonomous systems, and follow this up with modeling how attacks achieve their objectives. We argue that a large number of threats that can materialize as attacks and the costs of managing these attacks in cost effective ways require ranking threats using cyber threat modeling and cyber risk analysis techniques. The last segment of the paper describes a structured approach to mitigate high-risk threats.

Music Genre Classification Using Data Mining and Machine Learning

With accelerated advances in internet technologies users make listen to a staggering amount of multimedia data available worldwide. Musical genres are descriptions that are used to characterize music in music stores, radio stations and now on the Internet. Music choices vary from person to person, even within the same geographical culture. Presently Apple’s iTunes and Napster classify the genre of each song with the help of the listener, thus manually. We propose to develop an automatic genre classification technique for jazz, metal, pop and classical using neural networks using supervised training which will have high accuracy, efficiency and reliability, and can be used in media production house, radio stations etc. for a bulk categorization of music content.

Business Intelligence Dashboard Application for Insurance Cross Selling

Insurance Companies use Business Intelligence (BI) and Business Analytics (BA) to quantify their business and to predict their growth with the help of BI solutions. The primary objective of this paper is to build a software solution which provides a platform for insurance companies and ecommerce to find a set of tools and solutions that can be implemented for their business data analytics. The BI Dashboard application can be used by insurance companies to implement the concept of Cross-Selling and Up-selling of insurance products to their customers. The Ecommerce web based application is used to implement the concept of group-based collaborative marketing of products which internally uses data mining and clustering algorithms.

Reducing the data communication delay in wireless sensor networks

We study the problem of delay efficient scheduling of data communicated by sensor nodes for sensor fusion in wireless sensor networks. We model the sensor network as a tree in which sensor nodes collaboratively observe an event and transmit their measurements to intermediate sensor nodes that lie along the path the root of the tree. The root node functions as the sensor data fusion center that is responsible for aggregating distributed measurements, while internal tree nodes perform dual functionality - as local routers and as intermediate data fusion nodes. We are interested in two problems - one for minimizing the sum of the end-to-end delays of sensors, and the other for minimizing the maximum end-to-end delay of a sensor. Since these problems are computationally hard, we approach these problems by minimizing the average delay and the maximum delay one hop at a time. We provide low complexity, distributed optimal solutions for both these problems. Further, we show through simulations that by minimizing the delay hop by hop, we can achieve good delay performance relative to the global problem of minimizing the sum of the end-to-end delays.

Optimal maximum likelihood estimates fusion in distributed network of sensors

A distributed network of sensors leverages its performance by aggregating information gathered by individual sensors through the process of sensor data fusion. Estimating parameters using a centralized scheme entails transporting data from multiple sensors to a centralized fusion center, leading to high network bandwidth consumption. Additionally, fusing raw sensor data from sensors with different sensing modalities may not be feasible. We propose an alternative approach in which each sensor first individually estimates the unknown parameters based solely on its own sensor data. Since sensors may not have a-priori knowledge of the probability distribution of the unknown parameters, each sensor independently computes its individual maximum likelihood estimates. Individual estimates along with their sufficient statistics are then communicated to the fusion center, which treats these estimates as observations to compute the optimum aggregated maximum likelihood estimates by maximizing the new likelihood function of these observations. The proposed technique offers two significant advantages: (i) Since each sensor computes its individual estimates based solely on its own sensed data, it is easily applicable to sensor networks having multi-modal sensors, and (ii) As compared to raw sensor data, communicating estimates and their sufficient statistics to the fusion center requires substantially less network bandwidth. Performance of the aggregated estimates is evaluated through simulations and by computing the Cramer-Rao lower bound.