Search

Deep reinforcement learning has been successful in training an agent to play at human-level in Atari games. Here, outputs of the game images were fed into a deep neural network to compute optimal actions. Conceptually, reinforcement learning can be viewed as the intersection of planning, uncertainty, and learning. In this paper, deep reinforcement learning method is applied to solve a problem formulated as a partially observable Markov decision process (POMDP). Specifically, the input images are perturbed to introduce imperfect knowledge. POMDP formulations assume uncertainties in the true state space and thus a more accurate representation of the real-world scenarios. The deep Q-network is adopted to see if an optimal sequence of actions can be learned when the inputs are not fully observable. Experimental results indicated that optimal strategies were discovered by deep reinforcement learning in majority of test cases, albeit slower to converge to the optimal solution.

We describe a class of quantum algorithms to generate models of propositional logic with equal probability. We consider quantum stochastic flows that are the quantum analogues of classical Markov chains and establish a relation between fixed points on the two flows. We construct chains inspired by von Neumann algorithms using uniform measures as fixed points to construct the corresponding irreversible quantum stochastic flows. We formulate sampling models of propositions in the framework of adiabatic quantum computing and solve the underlying satisfiability instances. Satisfiability formulation is an important and successful technique in modeling the decision theoretic problems in a classical context. We discuss some features of the proposed algorithms tested on an existing quantum annealer D-Wave II extending the simulation of decision theoretic problems to a quantum context.

To investigate the origin and functional role of the brain's electrical activity in specific frequency ranges, scientists have employed biologically accurate computational models of the brain. The onset of extreme-scale computing brings the promise of allowing neuroscientists the ability to examine the brain in silico at unprecedented fidelity and resolution that is not possible using traditional methods such as the electroencephalogram. The large amount of data produced by high fidelity simulations are becoming increasingly difficult to save and transform into scientific insights for runtime simulations. One attractive solution is to modify the data analysis workflow from one that is accomplished post-simulation to one that is completed in situ at the cost of confining the analysis to the level of local data rather than a single global view. The missing global view of data has the potential to introduce inaccuracies in the simulation's findings. In this paper, we present the integration of in situ analysis into a simulation of the electrical activity of a brain's neuronal network to enable computational steering of the simulation itself at runtime (i.e., without stopping, performing post-simulation analysis, and restarting the simulation). We evaluate the accuracy of our in situ analysis versus the traditional post-simulation analysis, showing how we can gain meaningful global insights from local data. We demonstrate the integration's effectiveness in steering the electrical activity of a simulated neuronal network at runtime from electroencephalograph's (EEG) alpha frequency band (8-13Hz) to the beta frequency band (13-40Hz).

One significant challenge in neuroscience is understanding the cooperative behavior of large numbers of neurons. Models of neuronal networks allow scientists to explore the impact of differential neuronal connectivity using analysis techniques and information not available experimentally. However, modeling realistic neurobiological processes and encoding them in computer simulations is challenging, as increasing computing and data requirements are all of concern. In this work we study the performance of neocortex simulations using GEneral NEural SImulation System (GENESIS), a well-known multi-function brain simulation package, supported by high performance computing resources. The contribution of our work is threefold. First, we study the impact of platforms (i.e., single fat nodes versus high-end clusters) and their features on the performance and data generation for a small scale model of neocortex. Second, we assess the impact of the model complexity (i.e., number of cells and cell connectivity)on the performance and data generation for increasingly large versions of the neocortex model on high-end clusters. Third, we provide selected scientific results obtained by increasing the model complexity. We show that the more realistic and rigorous modeling of neocortex functions is computationally feasible but requires high performance computing resources to mitigate the growing computing and data requests of the associated simulations.

The optimization of injection gate locations in liquid composite molding processes by trial and error based methods is time consuming and requires an elevated level of intuition, even when high fidelity physics-based numerical models are available. Optimization based on continuous sensitivity equations (CSE) and gradient search algorithms focused towards minimizing the mold infusion time gives a robust approach that will converge to local optima based on the initial solution. Optimization via genetic algorithms (GA) utilizes natural selection as a means of finding the optimal solution in the global domain; the computed solution is at best, close to the global optimum with further refinement still possible. In this paper, we present a hybrid global–local search approach that combines evolutionary GAs with gradient-based searches via the CSE. The hybrid approach provides a global search with the GA for a predetermined amount of time and is subsequently further refined with a gradient-based search via the CSE. In our hybrid method, we utilize the efficiency of gradient searches combined with the robustness of the GA. The resulting combination has been demonstrated to provide better and more physically correct results than either method alone. The hybrid method provides optimal solutions more quickly than GA alone and more robustly than CSE based searches alone. A resin infusion quality parameter that measures the deviation from a near uniform mold volume infusion rate is defined. The effectiveness of the hybrid method with a modified objective function that includes both the infusion time and the defined mold infusion quality parameter is demonstrated.

This report acts as an in-depth guide to understanding the full feature set for visualizing the Ballistics Threat Geospatial Optimization code. This guide goes over the necessary steps to run the visualization and set up various scenarios to demonstrate the capabilities of the ballistics threat code. It also includes a detailed breakdown of how to interpret results from running the visualization.

The Black-Box Adaptive Fast Multipole Method (bbAFMM) has been generating some interest within the high-performance computing community as a tractable solution to the well-known n-body problem. The bbAFMM approximates the n-body solution using a series of independent functions or kernels that are attractive to high-performance code development using one or more graphics processor unit (GPU) devices. This work follows the analysis and implementation of the direct interaction called particle-to-particle kernel for a shared-memory single GPU device using the Compute Unified Device Architecture, revealing a performance boost of greater than 500 times over the corresponding serial central processing unit implementation. The objective of this work is to both document the implementation of the GPU kernel and provide a better understanding of the observed performance through an algorithmic analysis that focuses on arithmetic intensity, GPU memory bandwidth, GPU peak performance, and the defined Peripheral Component Interconnect Express bandwidth.

The Legion programming system from Stanford University was developed to address the specific needs of portable parallel heterogeneous programming with emphasis on program data. Data-centric Legion abstracts the underlying hardware such that the focus is algorithmic development rather than the idiosyncrasies of different computing architectures. Legion is part of a small but growing movement to treat parallel programming design and development in a task-oriented fashion and, as of this writing, is the only one to address this paradigm dynamically. This work employs a gravitational n-body simulation as a vehicle to introduce Legion to a wider audience of parallel programmers.

The High Performance Fortran (HPF) language and the Message Passing Interface (MPI) are two widely used methods to achieve parallelism on today's clusters and multiprocessor supercomputers. HPF is a distinct language providing extensions to Fortran 90/95 to express parallel execution paths and regions. MPI is a library of communication calls that can be inserted into modern high-level languages (C and Fortran). This paper discusses the use of the two approaches in a parallel finite element application for liquid composite manufacturing process modeling. The unstructured nature of the code provides an excellent opportunity to test both the computation and communication effectiveness of the two approaches. We discuss performance results based on implementations conducted on a modern massively parallel computing platform with a highly tuned processor interconnection network.

New computing architectures raise questions of technical challenges and investment risks to the high-performance computer (HPC) community. Difficulties exist in determining runtime performance ahead of time and lowering the high costs associated with code migration. Furthermore, the popularity of heterogeneous systems has added complexity in assessing resource capability. A mechanism to quickly and accurately predict performance would be instrumental in assessing the risk of HPC center procurements. A sound metric for performance prediction will help to determine applicability and feasibility of new architectures. To realize the art of anticipating performance, we surveyed the literature on benchmarking suites. This report describes the central characteristics a benchmarking suite should exhibit and derives a list of pertinent requirements for diverse computational resources at the US Army Research Laboratory. Modern benchmark suites are analyzed with respect to the derived list of requirements. This study leads to a benchmark candidate that can be leveraged for a state-of-the-art and next-generation supercomputing facility.

Under conditions requiring line-of-sight (LOS) communications, military units can benefit from a relaying system to connect nodes that do not have LOS to each other. An autonomous vehicle capable of positioning itself at an appropriate location, or moving between multiple locations as needed to maintain communications, can provide this capability while reducing the risk to personnel. If such a vehicle has sufficient capacity, it can also provide additional services such as local high-performance computing in an environment where these services would otherwise be unavailable. In order to move to the correct destination as quickly as possible, an off road navigation system should consider the constraints imposed by vehicle dynamics, because the fastest feasible path may have little similarity to the shortest possible path. This paper presents a method of determining a path with minimal travel time, and analyzes the effects that knowledge about the terrain has on the results of this algorithm.

The tactical edge (battlefields, disaster relief operations, etc.) is an ideal domain for utilizing mobile devices for computation. However, these situations require jobs to complete in a timely manner. Efficiently offloading computation can ensure timely job completion, but access to cloud computing infrastructure is typically limited, unreliable, or unavailable. As an alternative, we propose using vehicle-mounted high performance computers, called Tactical HPCs (T-HPCs), as offload targets. A common scenario includes multiple offload targets with different computation capabilities, multiple clients wanting to offload computations, and wireless ad-hoc connectivity. While offloading strategies exist, currently none address the requirements of task scheduling in this setting. This paper presents an algorithm for offloading tasks from multiple sources while taking into account the computation capabilities and current utilization of available offload targets. In addition, distributed scheduling avoids a single point of failure which could be disastrous on the tactical edge. We report the results of simulations showing that our method reduces job completion time compared to alternative offloading schemes.

An observation in supercomputing in the past decade illustrates the transition of pervasive commodity products being integrated with the world's fastest system. Given today's exploding popularity of mobile devices, we investigate the possibilities for high performance mobile computing. Because parallel processing on mobile devices will be the key element in developing a mobile and computationally powerful system, this study was designed to assess the computational capability of a GPU on a low-power, ARM-based mobile device. The methodology for executing computationally intensive benchmarks on a handheld mobile GPU is presented, including the practical aspects of working with the existing Android-based software stack and leveraging the OpenCL-based parallel programming model. The empirical results provide the performance of an OpenCL N-body benchmark and an auto-tuning kernel parameterization strategy. The achieved computational performance of the low-power mobile Adreno GPU is compared with a quad-core ARM, an x86 Intel processor, and a discrete AMD GPU.

This report documents the installation, acceptance testing, vendor interactions, and final system acceptance for production usage of a dedicated high performance computing project investment that was installed at the U.S. Army Research Laboratory facilities on the Aberdeen Proving Ground in MD. The uniqueness of this system lies in the use of the low latency 10 GigE (gigabit Ethernet) network fabric, many-core AMD Opteron central processing units (CPUs), Nvidia Fermi graphics processing unties (GPUs), and a mixture of CPU-only and CPU/GPU nodes. The final accepted production system is stable and capable of nearly 1 PetaFLOPS (1012 floating-point operations per second) of single precision performance.