This work demonstrates the autonomous command of a trained search canine to multiple waypoints using a novel state machine control algorithm. A hardware system is utilized in order to interface with the Global Position Satellite (GPS) system and with a tone and vibration generator for the purpose of accurately navigating and commanding the canine. An operational control algorithm for autonomous guidance of the canine is described in detail. Empirical results of an autonomously commanded canine are demonstrated with an 73% mission success rate for simple paths and a 62% mission success rate for complex paths. This work demonstrates a novel way to expand the capabilities of canines in a wide variety of missions, including search and detection.

This paper demonstrates a capability to use a developed embedded sensor suite to consistently track the position, motion behavior, and orientation of a canine. Quantifying and recording canine position and motion in real time provides a useful mechanism for objective analysis of canine trials and missions. We provide a detailed description of the sensor equipment, including the global position satellite (GPS) receiver and antenna, accelerometers, gyroscopes, and magnetometers. Sensors beyond GPS provide for higher frequency readings, a tolerance to GPS loss, and the ability to characterize canine orientation. We demonstrate integrating sensor measurements using an Extended Kalman Filter (EKF) to estimate the canine position and velocity during temporary GPS loss. The system supports the remote actuation of tone and vibration commands and reports commands in real time alongside sensor data. This extends the range at which a handler could monitor a canine and allows enhanced trial analysis using raw sensor data and visualizations. To illustrate the system capabilities, we performed a case study in the remote command and navigation of a trained canine by a professional trainer. The results of this case study are analyzed in terms of canine trial success, motion behavior analysis, and in the context of simulated GPS losses. We discuss other potential applications of the system in autonomous canine command, canine motion analysis, and non-canine applications.

This work seeks to contribute to software development education by motivating the use of engaging in-class and laboratory assignments. Ideally, these assignments should involve considerable student buy-in and should also evolve throughout the course to mimic real-world software development. Prior research is discussed, as well as several specific examples from two introductory programming classes. The ultimate contribution is a convincing argument to spend the extra effort to design better student projects.

This paper demonstrates an evolutionary algorithm for the optimization of an autonomous control algorithm for a trained canine. Autonomous guidance is relevant because use of canines, though beneficial in many applications, is limited by the necessity of close human supervision. A rules-based expert system using GPS data was initially developed for this purpose. This rules-based system is not without limitations. Primarily, it takes a significant investment of trainer and developer time to derive appropriate values to use for control of the canine. A multi-objective fitness metric was developed to optimize for important parts of the control algorithm, and parameters of the algorithm were optimized using evolutionary computation. In simulations the evolved parameters fit the data better than the hand-tuned parameters, and preliminary field trials showed a 67% mission success rate, which shows the feasibility of evolving parameters for the control algorithm.

The purpose of this work is to contribute to the reverse engineering of software by providing a means for identifying the type of compiler used to compile a Java class or Linux ELF file. A software framework is presented for extracting potentially useful information from class files and analyzing that information to classify future files. A General Regression Neural Network is implemented and optimized using evolutionary computation. In experimental results, the system can classify compiler type on an file it has not seen before with over 98% accuracy.

This dissertation demonstrates the autonomous command and navigation of a trained canine to multiple waypoints. A system is described consisting of a canine that can be guided autonomously to a number of waypoints by an automatic software control algorithm. A hardware system has been developed in order to interface with GPS, accelerometers, gyroscopes, magnetometers, and tone and vibration generators for the purpose of accurately commanding and dictating the motion, path, and commands given to a canine. A canine has been trained to effectively follow audio and vibration commands for guidance with a high degree of accuracy (71% mission success for simple paths and 63% mission success for complex paths). Both a Neural Networks approach and a State Machine Based approach to canine anomaly detection are presented, as well as strategies for anomaly correction. An operational control algorithm for autonomous guidance of the canine is described in detail. Finally, empirical results of an autonomously commanded canine are demonstrated with an 73% mission success rate for simple paths and a 62% mission success rate for complex paths.

This paper attempts to automate and replace human guidance in the control of a K-9 unit by modeling that guidance from observation. The ultimate research goal seeks to contribute toward the autonomous command of a trained K-9 unit by analyzing the movement and the behavior of the dog as it responds to command tones. Specifically, GPS and command signal information (from a human trainer) is recorded as a canine follows (or fails to follow) instructions as it moves toward a destination. The data is then processed into training instances and used as training data for a general regression neural network (GRNN). Then, the network is used to classify previously unseen test instances to determine if the behavior at that moment is normal or anomalous (in need of correcting tones). Both representation of training instances and the system parameters of the GRNN are optimized using a simple evolutionary hill-climber (EHC). Given even fairly limited initial data for training, the system performs well, producing relatively few false positives and false negatives in classification.

As library-based techniques of virus and intrusion detection prove insufficient to be means of protection for computer systems and networks, alternate methods must be incorporated into computer defense. This paper surveys a variety of Artificial Intelligence methods aimed at improving the effectiveness of modern computer security in the face of more prevalent, sophisticated, and ever-changing threats. Specifically, Artificial Immune Systems, Genetic Algorithms, Genetic Programming, and Neural Networks are discussed primarily within the auspices of threat analysis and system testing. Several implemented solutions are discussed and analyzed. Additionally, the paper provides suggestions for the synthesis of varying methods with traditional security in order to create a stronger, more robust security system.

The purpose of the Meta-Parallel Evolutionary System (MPES) is to develop fast, efficient parallel evolutionary systems for function optimization. Given an optimization problem and a set number of nodes available for the computation, the MPES searches for a strong, potentially heterogeneous combination of evolutionary algorithms to coordinate in order to effectively solve a problem. The Evolutionary Algorithms that are utilized in the parallel system are a Particle Swarm Optimizer (PSO), a variety of Genetic Algorithms (GAs), and an Evolutionary Hill-Climber Algorithm (EHC). The subpopulations communicate with each other via one or more centralized buffers. At a higher level exists the MPES, which uses evolutionary methods in order to discover parameters for effective parallel systems. This methodology provides an immediate benefit in the form of a strong tool to solve the optimization problem. Further, it provides a long-term benefit by identifying a system that has the potential to effectively solve other difficult optimization problems with a similar search space.

In this paper, we present four genetic protocols for solving randomly generated distributed constraint satisfaction problems. These genetic protocols are based on an evolutionary paradigm known as a society of hill-climbers (SoHC) and are thus referred to as genetic SoHCs (GSoHCs). The difference between the SoHC and the GSoHCs is that each of the GSoHCs use a distributed restricted recombination operator. We compare the SoHC and GSoHC protocols on a test suite of 400 randomly generated distributed constraint satisfaction problems (DisCSPs) that are composed of asymmetric constraints (referred to as DisACSPs). In a second experiment, we compare the best performing GSoHC and the SoHC on an additional 1100 randomly generated DisACSPs in order to compare their performances across the phase transition. Our results show that all of the GSoHCs dramatically outperform the SoHC protocol even at the phase transition where, on average, the most difficult DisACSPs reside.

Evolutionary Computation (EC) is the field of study devoted to problem solving using simulated evolution. In this paper, evolutionary operators are applied and a number of societies of hill-climbers (SoHCs), such as a genetic SoHC (GSoHC) and an evolutionary SoHC (ESoHC) are employed for solving randomly generated distributed asymmetric constraint satisfaction problems (DisACSPs). Further, we develop an Estimation of Distribution Algorithm SoHC (EDA-SoHC) variant using a uniform mutation operator. This variant produces offspring by drawing genetic material from a distribution of above-average individuals in the population. In this paper, we compare GSoHCs using distributed restricted forms of single-point, two-point, modified two-point, and uniform crossover. The GSoHCs are also compared with an ESoHC that uses a distributed restricted form of uniform mutation and a simple SoHC which does not use any evolutionary operators. Finally, we compare the SoHC, GSoHCs, and ESoHC to the EDA-SoHC.

The nuclear electric vehicle optimization toolset (NEVOT) optimizes the design of all major nuclear electric propulsion (NEP) vehicle subsystems for a defined mission within constraints and optimization parameters chosen by a user. The tool currently uses a number of evolutionary computations (ECs) for designing NEP vehicles. Since evaluating candidate vehicle designs is computationally expensive, it is important that a set of robust control parameters be discovered. In order to accomplish this, a meta-genetic algorithm (meta-GA) was developed to discover control parameters for generational, steady-state, and steady-generational GAs as well as for particle swarm optimizers (PSOs) with ring, star, and random topologies. Our results show that the high-performance GAs are more efficient than the high-performance PSOs on a NASA asteroid mission problem.

The paper illustrates the motivation for moving away from a paper-based ticket writing system to digital system using a portable devise. A simple to use application with a graphical user interface was developed for The Microsoft Smartphone.

In this paper, we present simple and genetic forms of an evolutionary paradigm known as a society of hill-climbers (SoHC). We compare these simple and genetic SoHCs on a test suite of 400 randomly generated distributed constraint satisfaction problems (DisCSPs) that are composed of asymmetric constraints (referred to as DisACSPs). Our results show that all of the genetic SoHCs dramatically outperform the simple SoHC even at the phase transition where the most difficult DisACSPs reside.

In this paper we explore using several different learning algorithms in order to classify whether or not weather conditions were appropriate for tennis playing. Specifically, we created a Naïve Bayesian Classifier (NBC), an Artificial Immune System (AIS), and a General Regression Neural Network (GRNN). Further, for the GRNNs, we used a Particle Swarm Optimizer (PSO) and an Estimation of Distribution Algorithm (EDA) in order to find the optimal values of sigma. Each algorithm and its corresponding parameter settings is described. The algorithms were evaluated based on both training set error (sum-squared error) and test set error (percentage of correctly identified instances). The results of these experiments are presented and some rationale is given.

In this paper we explore using several different evolution strategies (ES) in order to optimize the parameters for the F6 function for minimization. Specifically, we created a Simple (1 + 1)-ES (SES), a (µ + 1)-ES (MPOES), a (µ + λ)- ES (MPLES), and a (µ, λ)-ES (MCLES). Further, we explored performance-enhancing modifications to each of these algorithms. Each algorithm and its corresponding modifications and parameter settings is described. The algorithms were tested using a direct mapping between the output of the F6 function and the fitness function for the ES. The results of those experiments are presented and some rationale is given.