Search

Highway on-ramp merge junctions remain a major bottleneck in transportation networks. However, with the introduction of Connected Autonomous Vehicles (CAVs) with advanced sensing and communication capabilities modern algorithms can capitalize on the cooperation between vehicles. This paper enhances highway merging efficiency by optimally coordinating CAVs in order to maximize the flow of vehicles while satisfying all safety constraints. Focus is also placed on the effect of varying priorities of different vehicle classes in selecting the best merging sequence. Our algorithm is capable of real time operation through parallel computation, optimized merge sequence generation and management of the diverse needs of heterogeneous (multi-class) traffic. Results are verified through a realistic traffic simulation software.

Cyber threat information can be utilized to investigate incidents by leveraging threat-related knowledge from prior incidents with digital forensic techniques and tools. However, the actionability of cyber threat information in digital forensics has not yet been evaluated. Such evaluation is important to ascertain that cyber threat information is as actionable as it can be and to reveal areas of improvement. In this study, a dataset of cyber threat information products was created from well-known cyber threat information sources and its actionability in digital forensics was evaluated. The evaluation results showed a high level of cyber threat information actionability that still needs enhancements in supporting some widely present types of attacks. To further enhance the provision of actionable cyber threat information, the development of the new TREVItoSTIX Autopsy module is presented. TREVItoSTIX allows the expression of the findings of an incident investigation in the structured threat information expression format in order to be easily shared and reused in future digital forensics investigations.

We propose an adaptive state aggregation scheme to be used along with temporal-difference reinforcement learning and value function approximation algorithms. The resulting algorithm constitutes a two-timescale stochastic approximation algorithm with: (a) a fast component that executes a temporal-difference reinforcement learning algorithm, and (b) a slow component, based on online vector quantization, that adaptively partitions the state space of a Markov Decision Process according to an appropriately defined dissimilarity measure. We study the convergence of the proposed methodology using Bregman Divergences as dissimilarity measures that can increase the efficiency and reduce the computational complexity of vector quantization algorithms. Finally, we quantify its performance on the Cart-pole (inverted pendulum) optimal control problem using Q-learning with adaptive state aggregation based on the Self-Organizing Map (SOM) algorithm.

Teachers often spend a long time assessing work and examinations for students; automatic short answer grading dramatically reduces any time spent grading. Also is can be tough for teachers to be entirely unbiased when assessing work. Towards addressing this, a short answer automatic grading agent utilizing answer segmentation for GCSE Computer Science with semantic matching was designed and implemented using a four stage pipeline. The short answer marking agent has over a 96% accuracy rating and marks over 55% of student answers with the rest being referred for human grading. The human graded answers are then used to further automatically improve the system. In the future the agent could easily be extended to other subjects and used for staff training in a wide variety of areas or possibly even to form the basis for active tutoring systems.

We propose a family of parametric interaction functions in the general Cucker-Smale model such that the mean-field macroscopic system of equations can be iteratively solved in an optimization scheme aiming to learn the inter-action dynamics of the microscopic model from observations of macroscopic quantities. We treat the interaction functions as Green’s functions for a semi-linear Poisson differential operator, which allows the transformation of the non-local interaction terms of the macroscopic model into a system of PDEs. The resulting system of hydrodynamic equations is efficiently solved as part of an iterative learning algorithm that estimates the interaction function from particle density evolution data. Finally, we utilize the proposed interaction function model to formulate an efficient learning algorithm based on observations from particle trajectories, and discuss the trade-offs associated with each approach.

GPUs in datacenters and cloud environments are mainly offered in a dedicated manner to applications, which leads to GPU under-utilization. Previous work has focused on increasing utilization by sharing GPUs across batch and user-facing tasks. With the presence of long-running tasks, scheduling approaches without GPU preemption fail to meet the SLA of user-facing tasks. Existing GPU preemption mechanisms introduce variable delays up to several seconds, which is intolerable, or require kernel source code, which is not always available.In this paper, we design TReM, a GPU revocation mechanism that stops a task at any point in its execution. TReM has a constant latency, of about 5ms to stop the currently executing kernel and about 17ms to start a new task. TReM does not store the state of the revoked kernel to obviate transfer latencies. We design and implement two scheduling policies, Priority and Elastic, that prioritize user-facing over batch tasks and utilize TReM to improve SLAs for user-facing tasks. To evaluate TReM, we use a workload generator that creates workloads with different characteristics, based on real traces. TReM reduces SLA violations by up to 10% compared to baseline policies that do not use a revocation mechanism. TReM incurs negligible overhead for non-revoked tasks and wastes only 3% of computation due to revocations for the workloads we examine.

We propose a human-robot interface based on potentials recorded through surface Electroencephalographic sensors, aiming to decode human visual attention into motion in three-dimensional space. Low-frequency components are extracted and processed in real time, and subspace system identification methods are used to derive the optimal, in mean squared sense, linear dynamics generating the position vectors. This results in a human-robot interface that can be used directly in robot teleoperation or as part of a shared-control robotic manipulation scheme, feels natural to the user, and is appropriate for upper extremity amputees, since it requires no limb movement. We validate our methodology by teleoperating a redundant, anthropomorphic robotic arm in real time. The system's performance outruns similar EMG-based systems, and shows low long-term model drift, indicating no need for frequent model re-training.

We propose a framework based on port-Hamiltonian modeling formalism aimed at learning interaction models between particles (or networked systems) and dynamical properties such as trajectory symmetries and conservation laws of the ensemble (or swarm). The learning process is based on approaches and platforms used for large scale optimization and uses features such as automatic differentiation to compute gradients of optimization loss functions. We showcase our approach on the Cucker-Smale particle interaction model, which is first represented in a port-Hamiltonian form, and for which we re-discover the interaction model, and learn dynamical properties that are previously proved analytically. Our approach has the potential for discovering novel particle cooperation rules that can be extracted and used in cooperative control system applications.

Software defined systems use virtualization technologies to provide an abstraction of the hardware infrastructure at different layers. Ultimately, the adoption of software defined systems in all cloud infrastructure components will lead to Software Defined Cloud Computing. Nevertheless, virtualization has already been used for years and is a key element of cloud computing. Traditionally, virtual machines are deployed in cloud infrastructure and used to execute applications on common operating systems. New lightweight virtualization technologies, such as containers and unikernels, appeared later to improve resource efficiency and facilitate the decomposition of big monolithic applications into multiple, smaller services. In this work, we present and empirically evaluate four popular unikernel technologies, Docker containers and Docker LinuxKit. We deployed containers both on bare metal and on virtual machines. To fairly evaluate their performance, we created similar applications for unikernels and containers. Additionally, we deployed full-fledged database applications ported on both virtualization technologies. Although in bibliography there are a few studies which compare unikernels and containers, in our study for the first time, we provide a comprehensive performance evaluation of clean-slate and legacy unikernels, Docker containers and Docker LinuxKit.

Cyber security education is an emerging field facing many challenges, while demand in skilled cyber security personnel is rising. Serious games and game-based learning approaches constitute a promising opportunity for learning and training in cyber security. However, as game-based learning is a new field, cyber security game-based approaches lack common methodologies and design standards that will provide useful interpretations of the possibilities of such approaches. In this light, we present the Conceptual Framework for e-Learning and Training (COFELET), a framework for the design and implementation of cyber security serious games. The COFELET framework promotes game-based approaches that envisage the improvement of cyber security education pedagogical effectiveness by embracing modern learning theories and innovative teaching approaches. Moreover, the COFELET framework foresees the implementation of a sustainable lifecycle of learning, updating and reinforcing knowledge and skills by adopting multidisciplinary approaches based on sound cyber security methodologies, models and strategies. The COFELET framework realizes cyber security as an attractive and accessible subject endorsing game-based learning and serious games, cyber world simulations, and hacking activities.

This paper addresses the problem of physical human-robot collaboration for object manipulation. In particular, we consider a human-robot architecture where the human has exclusive knowledge of the object’s desired trajectory and the robot tries to assist actively, by carrying the object’s load in order to reduce the human effort that is required to achieve the desired tracking behavior. The robot estimates the human’s desired motion via a prescribed performance estimation law that drives the estimation error to an arbitrarily small residual set. This estimation is further employed in the object dynamics equation to compute the interaction force between the human and the object. Subsequently, an impedance control scheme is designed based on the aforementioned estimations, achieving significant reduction on the required human effort, despite the uncertainty in the robot dynamics. The feedback relies exclusively on the robot’s force/torque, position as well as velocity measurements and no a priori explicit information on the task is required. Finally, extensive experimental results clarify the proposed method and verify its efficiency.

The paper presents neuroevolution approach to a crawler robot motion that autonomously solves sequences of navigation and flipper control tasks to overcome obstacles in 3D simulation domain. When modelling scenarios of robot locomotion, we used our model of a novel Russian crawler robot “Engineer” in ROS/Gazebo. The modelled robot measured obstacle's height by scanning a vertical profile of a terrain with 2D LIDAR and moved through 3D environment, adjusting its flippers to a relief by commands of the HyperNEAT neural network. As the result, our neuroevolution method was trained and tested on simulated data with a set of obstacles, demonstrating original solutions to robot navigation in 3D scene.

Biomimetic centipede robots can be well-suited to a number of applications, including search-and-rescue around demolished rubble, logistics in rocky and hazardous areas, and more. The design space for such robots in quite large, with numerous open possibilities for body and leg shapes, configurations and numbers of components. On the other hand, realistic systems aim toward optimizing various performance measures, some of which can be directly calculated from the design parameters, while others require physical experimentation and simulation. In this paper, we present parametric models of centipede robots as well as rough terrains, and then we evaluate the performance of different robot designs across different terrains. We incrementally derive better designs through an appropriately crafted evolutionary algorithm. Promising results as well as insights are discussed. Finally, we provide a clear list of future steps and a forward-looking conclusion.

This paper considers legged robots, in author's case Nao robot, and describes a method for Real-world large-step stair-climbing for small humanoids using Evolutionary Computation. We start by providing background on existing climbing robots and popular methods to make them move, and then discuss the problem of stair climbing robots. Of particular interest is a class of problems, when the robot is not able to pick up its leg on height of one stair of the ladder because of the limitations of its geometrical parameters. We provide a genetic algorithm solution for this task; explain design of the chromosome, forming of initial population, crossover and mutation operations. Results of experiments in the Webots simulator are described. Finally, we describe our experiments with the real robot, and discuss our results.

High dynamic capabilities of industrial robots make them dangerous for humans and environment. To reduce this factor and advance collaboration between human and manipulator fast and reliable collision detection algorithm is required. To overcome this problem, we present an approach allowing to detect collision, localize action point and classify collision nature. Internal joint torque and encoder measurements were used to determine potential collisions with the robot links. This work proposes two ways of solving this problem: using classical analytical approach and learning approach implemented with neural network. The suggested algorithms were examined on the industrial robotic arm Kuka iiwa LBR 14 R820, ground truth information on the contact nature and its location were obtained with 3D LIDAR and camera.

Cloud computing is a widespread and commonly used computing paradigm. Cloud relied on virtualization technology to be a scalable and efficient distributed system. In traditional cloud infrastructures, hypervisors like KVM create and manage Virtual Machines according to users’ demands. Recently a more lightweight virtualization technique is being rapidly adopted and it is based on containers.Although running a container on top of Virtual Machine is a common case, according to our research it has never been studied before how this affects the container’s performance. In this work, we conducted several experiments to study how the container performance is affected by the additional virtualization layer of the Virtual Machine. We used Docker to deploy and run our containers, KVM hypervisor for the Virtual Machines and we executed a series of well-known benchmarks on different operating systems.

This paper introduces a contemporary platform, called FGAWeb, that can be utilized in order to administrate flexible games. FGAWeb platform is focused on differentiating the content of puzzle and quiz games in order to make them more suitable for use in educational settings. Instructors can utilize this platform in order to create and edit game activities as well as, to have access to game and learning analytics. FGAWeb platform offers an API that makes the whole system accessible from any game engine that supports the HTTP protocol for data transfers.

Joystick-based teleoperation is a dominant method for remotely controlling various types of robots, such as excavators, cranes, and space telerobotics. Our ultimate goal is to create effective methods for training and assessing human operators of joystick-controlled robots. Towards that goal, an extensive study consisting of a total of 38 experimental subjects on both simulated as well as a physical robot, using either no feedback or auditory feedback, has been performed. In this paper, we present the complete experimental setup and we report only on the 18 experimental subjects teleoperating the simulated robot. Multiple observables were recorded, including not only joystick and robot angles and timings, but also subjective measures of difficulty, personality and usability data, and automated analysis of facial expressions and blink rate of the subjects. Our initial results indicate that: First, that the subjective difficulty of teleoperation with auditory feedback has smaller variance as compared to teleoperation without feedback. Second, that the subjective difficulty of a task is linearly related with the logarithm of task completion time. Third, we introduce two important indicators of operator performance, namely the Average Velocity of Robot Joints (AVRJ), and the Correct-to-Wrong-Joystick Direction Ratio (CWJR), and we show how these relate to accumulated user experience and with task time. We conclude with a forward-looking discussion including future steps.