Laurence D. Merkle

Measuring the effectiveness of robots in teaching computer science

We report the results of a year-long experiment in the use of robots to teach computer science. Our data set compares results from over 800 students on identical tests from both robotics and non-robotics based laboratory sessions. We also examine the effectiveness of robots in encouraging students to select computer science or computer engineering as a field of study.Our results are negative: test scores were lower in the robotics sections than in the non-robotics ones, nor did the use of robots have any measurable effect on students choice of discipline. We believe the most significant factor that accounts for this is the lack of a simulator for our robotics programming system. Students in robotics sections must run and debug their programs on robots during assigned lab times, and are therefore deprived of both reflective time and the rapid compile-run-debug cycle outside of class that is an important part of the learning process. We discuss this and other issues, and suggest directions for future work.

Simple genetic algorithm parameter selection for protein structure prediction

Selection of run-time parameters is a critical step in the application of genetic algorithms (GAs). Numerous investigations have discussed parameter set selection, both theoretically and empirically. Theoretical work has focused on the choice of population size, while empirical studies cover a wide range of GA parameters. Theory suggests population sizes which increase exponentially with string length. The available experimental data suggests small populations perform consistently well, but the test problems are limited to small string lengths. Thus, we still do not have a complete understanding of how parameters should be chosen, especially for problems with large string lengths. This study extends Schaffers (1989) results by performing a similar empirical analysis of GA parameters on a real-world application (protein structure prediction), with longer string lengths and a very large number of local optima. Relationships between population size, mutation rates and crossover rates similar to those reported by Schaffer are shown.

Automated network forensics

The purpose of this research is to investigate the automated analysis of network based evidence in response to cyberspace attacks. The automated analysis techniques to be developed and studied will combine the efficiency of both existing and novel local search techniques with the scalability and robustness of evolutionary computation and other computational intelligence techniques.

Hybrid genetic algorithms for minimization of a polypeptide specific energy model

A hybrid genetic algorithm for polypeptide structure prediction is proposed which incorporates efficient gradient-based minimization directly in the fitness evaluation. Fitness is based on a polypeptide specific potential energy model. The algorithm includes a replacement frequency parameter which specifies the probability with which an individual is replaced by its minimized counterpart. Thus, the algorithm can implement either Baldwinian, Lamarckian, or probabilistically Lamarckian evolution. Experiments are described which compare the effectiveness of the genetic algorithm with and without the local minimization operator, and for various probabilities of replacement. The experiments apply the techniques to the minimization of the ECEPP/2 energy model for [Met] Enkephalin. Using fitness proportionate selection, the hybrid approaches obtain better energies (and better basins of attraction) than the standard genetic algorithm, and often find the global minimum. When tournament selection is used, the results are qualitatively similar, except that the hybrid approaches are prone to premature convergence.

Hybrid genetic algorithms for polypeptide energy minimization

Efforts to predict polypeptide structures nearly always assume that the native conformation corresponds to the global minimum free energy state of the system. Given this assumption, a necessary step in solving the problem is the development of efficient global energy minimization techniques. We describe a hybrid genetic algorithm which incorporates efficient gradient-based minimization directly in the fitness evaluation, which is based on a general full-atom potential energy model. The algorithm includes a replacement frequency parameter which specifies the probability with which an individual is replaced by its minimized counterpart. Thus, the algorithm can implement either Baldwinian, Lamarckian, or probabilistically Lamarckian evolution. We also describe experiments comparing the effectiveness of the genetic algorithm with and without the local minimization operator, with various probabilities of replacement. The experiments apply the techniques to the minimization of the CHARMM potential for [Met]Enkephalin. When fitness proportionate selection is used, the Baldwinian, Lamarckian, and probabilistically Lamarckian approaches obtain better energies (and better basins of attraction) than the standard genetic algorithm. This suggests that the low-energy local minima in polypeptide energy landscapes occur sufficiently regularly to benefit from the proposed hybrid approaches. When tournament selection is used, the results are qualitatively similar, except that the hybrid approaches are prone to premature convergence. Increasing replacement frequency reduces the tendency toward premature convergence for the experiments performed here.

Polypeptide structure prediction: real-value versus binary hybrid genetic algorithms

A b s t r a c t Energy minimization efforts to predict polypeptide structures assuule their native conformation corresponds to the global minimum free energy state. Given this assumption, the problem becomes that of developing efficient global optinfization techniques applicable to polypeptide energy models. This general structure prediction objective is also known as the protein folding problem. Our prediction algorithms, based on general fifil-atom potential energy models, are expanded to incorporate domain knowledge into the search process. Specifically. we evaluate the effectiveness of a real-valued genetic algorithm exploiting domain knowledge about certain dihedral angle values inorder to limit the search space. We contrast this approach with our hybrid binary genetic algorithms. Various experiments apply these techniques to nfinimization of the potential energy for the specific proteins [Met]-Enkephalin and Polyala-nine using the CHARMM energy model.

VIRTUAL PROTOTYPING OF MICROWAVE DEVICES USING MHD, PIC, AND CEM CODES

The Directed Energy Directorate of the Air Force Research Laboratory has the Air Force Responsibility for the development of high power microwave (HPM) weapons. HPM devices tend to fall within the category of ultrawideband or narrowband, each having advantages and difficulties in development and application. The Center for Plasma Theory and Computation has developed a suite of scientific software to aid in the development of the components of such devices. A typical narrowband high power microwave device is made up of 3 components: a source of pulsed power which releases stored energy in the form of a fast (~nsec to msec) applied voltage, a beam/cavity interaction region in which the kinetic energy of a beam is transformed to microwave radiation, and an antenna which is used to direct the microwave radiation. These components may be categorized by their density of charged particles; the pulsed power devices often involve high-density plasmas, the beam/cavity sources have a low density of charged particles, and it is desired that the antennas have no charged particles. These regimes of charged particle density are most efficiently simulated with magnetohydrodynamic (MHD), particle-in-cell (PIC), and computational electromagnetic (CEM) techniques, respectively.

Game Development for Computer Science Education

Games can be a valuable tool for enriching computer science education, since they can facilitate a number of conditions that promote learning: student motivation, active learning, adaptivity, collaboration, and simulation. Additionally, they provide the instructor the ability to collect learning metrics with relative ease. As part of 21st Annual Conference on Innovation and Technology in Computer Science Education (ITiCSE 2016), the Game Development for Computer Science Education working group convened to examine the current role games play in computer science (CS) education, including where and how they fit into CS education. Based on reviews of literature, academic research, professional practice, and a comprehensive list of games for computing education, we present this working group report. This report provides a summary of existing digital games designed to enrich computing education, an index of where these games may fit into a teaching paradigm using the ACM/IEEE Computer Science Curricula 2013 [13], and a guide to developing digital games designed to teach knowledge, skills, and attitudes related to computer science.

Autonomous agents as air combat simulation adversaries

An autonomous agent is a device (or being) that can cope effectively in a defined environment, despite the fact that the environment is not completely controlled or even certainly known by the agent. If an autonomous agent is to be robust in such an environment, it must be able to plan its activities, react quickly to unforeseen events, and execute planned or modified behaviors to achieve goals. Building such agents is not trivial, but can yield high payoffs in environments such as air combat. We have begun to develop autonomous agents which exhibit appropriate behaviors for simulated air combat. Simulated air combat is a forgiving environment that allows us to study issues in building autonomous agents, but also meets real Air Force needs for training. In this paper, we outline the requirements of autonomous adversaries for air combat simulation and describe two prototype systems we have developed to lay the foundation for a multi-faceted research thrust in autonomous agent technology.

Metaoptimization of the in-lining priority function for a compiler targeting a polymorphous computing architecture

Leading polymorphous computing architecture (PCA) efforts include the Raw Architecture Workstation (Raw) and the Tera-op Reliable and Intelligently Adaptive Processing System (TRIPS), both of which are tile-based. The Raw toolchain places responsibility for program decomposition on the programmer, but the TRIPS toolchain automatically generates hyperblocks and allocates them to processing elements. This report identifies evolutionary computation (EC) techniques that enable and that are enabled by PCA technology, focusing on application of EC in enhancing the effectiveness of the TRIPS toolchain, including the Scale compiler. In particular, computational experiments are described that investigate the application of genetic programming to the meta-optimization of the priority function used to increase the number of instructions per hyperblock in the in-lining optimization phase of Scale. Results suggest continued experimentation with larger population sizes and more generations.