State-of-the-art image classification systems utilize powerful machine-learning-based tools such as Convolutional Neural Networks (CNNs). These networks can achieve high recognition accuracies, but suffer from a black-box problem where the inner workings are incomprehensible by humans that seek to use them. In this paper, a Feature-Oriented Cellular Automata Learning (FOCAL) system is developed to extend traditional gradient-filter-based methods by implementing a Cellular Automata (CA) reasoner utilizing rule-based primitives for determining mutual agreement between neighboring pixels. This novel method is demonstrated to identify features more accurately than standard filter methods and produce classification results that are competitive with typical CNNs, while also allowing a-priori definition of important features facilitating explainable feature classification decision processes. Experiments spanning a variety of influential factors indicate that rebaselining and normalization are vital to the success of the CA-based approach. Furthermore, within certain models, the use of CA is shown to reduce computational demand by over 90% while incurring only a 2% reduction in classification accuracy. Finally, the scalability of the FOCAL system is investigated using the CIFAR-10 dataset and contemporary Deep Neural Networks, and shown to encourage promising avenues of research into explainability while reducing computational processing demands.

This article describes our long-term research into automated story generation and our resulting story generation architecture called fAIble that incorporates several innovations. fAIble determines each event that occurs in the tale using a combination of scripted sequences and stochastically chosen events. The probability of an event occurring is based on the skills and personalities of the characters who have agency. Event selection is also influenced by the context of the situation faced by the characters. Each event is associated with a description in grammatically-correct natural language that can be narrated orally via text-to-speech. We describe the evolution of fAIble, its architecture and the results of our independent evaluation of each of the four progressively developed fAIble prototypes (fAIble 0, I, II and III), as tested with human test subjects. On a continuous scale where 0 means unacceptable, 1 means acceptable and 2 means optimal, the composite human test subject rating average from the independent tests of the prototypes was 0.933. The paper also describes a summative assessment where test subjects were asked to review stories from all four prototypes and rank them comparatively. These comparative results indicate an improvement from the original (fAIble 0) to the last one (fAIble III).

In this study, we investigate the effectiveness of using the PID (Proportional - Integral - Derivative) control loop factors for modifying response thresholds in a decentralized, non-communicating, threshold-based swarm. Each agent in our swarm has a set of four thresholds, each corresponding to a task the agent is capable of performing. The agent will act on a particular task if the stimulus is higher than its corresponding threshold. The ability to modify their thresholds allows the agents to specialize dynamically in response to task demands. Current approaches to dynamic thresholds typically use a learning and forgetting process to adjust thresholds. These methods are able to effectively specialize once, but can have difficulty re-specializing if the task demands change. Our approach, inspired by the PID control loop, alters the threshold values based on the current task demand value, the change in task demand, and the cumulative sum of previous task demands. We show that our PID-inspired method is scalable and outperforms fixed and current learning and forgetting response thresholds with non-changing, constant, and abrupt changes in task demand. This superior performance is due to the ability of our method to re-specialize repeatedly in response to changing task demands.

We introduce functionality-cognizant Genetic Algorithms (GAs) and graph-based operators to tackle the challenging search landscape of combinational digital circuit design. We introduce a novel circuit representation that builds upon Cartesian Genetic Programming (CGP), a popular grid-based method for representing directed graphs of connected components. Leveraging this, we introduce an original crossover operator that accounts for circuit component functionality and connectivity, as opposed to CGP, which only considers positional information in the chromosome. Additionally, we propose an innovative set of mutation operators and demonstrate successful evolution of fully functional and minimally sized common digital circuits including a variety of binary encoders and adders. Following successful synthesis of a four-bit adder, we present a generalizable machine learning approach for multi-layered search and optimization problems.

In this paper, we compare the performance of a canonical genetic algorithm (CGA), the Self Adaptive Genetic Algorithm (SAGA), and a feed-forward neural network (FFNN) on a predictive modeling problem with incomplete data. Predictive modeling involves learning relationships between the features and labels of the data points in a dataset. Datasets with missing input values may cause problems for some learning algorithms by biasing the learned models. Imputation refers to techniques for replacing missing data through methods such as statistical probabilities, multivariate analysis, machine learning, or K-nearest neighbors.

Virtual reality (VR) researchers have had a long-standing interest in studying locomotion for developing new techniques, improving upon prior ones, and analyzing their effects on users. To help organize prior work, several researchers have presented taxonomies for categorizing locomotion techniques in general. More recently, researchers have begun to conduct systematic reviews to better understand what locomotion techniques have been investigated. In this paper, we present our own systematic review of locomotion techniques based on a well-established taxonomy, and we use k-means clustering to identify to what extent locomotion techniques have been explored. Our results indicate that selection-based, walking-based, and steering-based locomotion techniques have been moderately to highly explored while manipulation-based and automated locomotion techniques have been less explored. We also present results on what types of metrics have been used to evaluate locomotion techniques. While usability, discomfort, and travel performance metrics have been moderately to highly explored, other metrics, such as biometrics, user experience, and emotions, have been less explored.

In this work, we investigate the application of a multi-objective genetic algorithm to the problem of task allocation in a self-organizing, decentralized, threshold-based swarm. We use a multi-objective genetic algorithm to evolve response thresholds for a simulated swarm engaged in dynamic task allocation problems: two-dimensional and three-dimensional collective tracking. We show that evolved thresholds not only outperform uniformly distributed thresholds and dynamic thresholds but achieve nearly optimal performance on a variety of tracking problem instances (target paths). More importantly, we demonstrate that thresholds evolved for some problem instances generalize to all other problem instances, eliminating the need to evolve new thresholds for each problem instance to be solved. We analyze the properties that allow these paths to serve asuniversal training instancesand show that they are quite natural. Aftera priorievolution, the response thresholds in our system are static. The problem instances solved by the swarms are highly dynamic, with schedules of task demands that change over time with significant differences in rate and magnitude of change. That the swarm is able to achieve nearly optimal results refutes the common assumption that a swarm must be dynamic to perform well in a dynamic environment.

Path planning, the task of finding an obstacle-avoiding, shortest-length route from source to destination is an interesting theoretical problem with numerous applications. We present an improved genetic algorithm for path planning in a continuous, largely unconstrained real-world environment. We introduce a new domain-specific crossover operator based on path intersections. We also implement a new path correction operator that eliminates obstacle collisions from a path, leading to a dramatic search improvement despite the conceptual simplicity of the correction. Finally, in place of a standard binary measure of obstacle collisions, we present a new optimization objective measuring the degree to which a path intersects obstacles. Due to these improvements, individually and in combination, our algorithm is able to solve scenarios that are considerably more complex and exist in a more general environment than those that appear in the literature. We demonstrate the utility of our system through testing onboard an autonomous micro aerial vehicle. Further, our approach demonstrates the utility of domain-specific genetic operators for path planning. We hypothesize that such operators may be beneficial in other domains.

The buildup of inaccuracies from frequent and imperfect data averaging can negatively impact system behavior. One potential culprit is improper orientation averaging, such as when combining data from multiple sensors, or reconciling preferences from multiple agents. In practice, the currently prevalent methodology of averaging 2D orientations is that of adding orientation vectors, which minimizes the Euclidean (or chord) distance among the vectors, instead of the geodesic (or arc) distance, resulting in inaccurate or even entirely incorrect averages. While an arc-minimizing alternative exists, it is only defined for angle averaging, posing an issue if orientations also possess a meaningful magnitude within the domain. In this work, we present an iterative weighted 2D orientation arc-based averaging algorithm that minimizes squared arc-lengths between points, incorporates orientation magnitudes as weights, and allows for multiple equally valid averages to be produced whenever applicable. We compare a vector sum approach and the weighted arc-based approach as applied to collaborative transport with obstacle avoidance, and showcase the behavioral advantages of the arc-based weighted averaging.

Nanoclusters are small clumps of atoms of one or several materials. A cluster possesses a unique set of material properties depending on its configuration (i.e. the number of atoms, their types, and their exact relative positioning). Finding and subsequently testing these configurations is of great interest to physicists in search of new advantageous material properties. To facilitate the discovery of ideal cluster configurations, we propose the Cluster Energy Optimizing GA (CEO-GA), which combines the strengths of Johnston's BCGA [18], Pereira's H-C&S crossover [25], and two new mutation operators: Local Spherical and Center of Mass Spherical. The advantage of CEO-GA is its ability to evolve optimally stable clusters (those with lowest potential energy) without relying on local optimization methods, as do other commonly used cluster evolving GAs, such as BCGA.

This work investigates the impact of inter-agent variation on the performance of a multi-agent system based control system. The control system is composed of a multi-agent team and aggregates the actions of the individual agents of the team to form a single output. In the process, this approach attempts to use error and variation that naturally occurs in physical systems to the system's advantage. We apply this multi-agent control system to a multiple task allocation problem and present an analytical model with which we can study the conditions under which such a system is expected to stabilize. We then compare the expected behavior of the analytical model with an empirical agent-based simulation applied to a tracking problem.

One of the expected benefits of multi-agent systems is their capacity for redundancy which is expected to promote system robustness. We examine the impact agents' probabilistic action has on generating redundancy in multi-agent systems with no direct inter-agent communication. We present a mathema

A recent theoretical investigation of modular representations shows that certain modularizations can introduce a distance bias into a landscape. This was a static analysis, and empirical investigations were used to connect formal results to performance. Here we replace this experimentation with an introductory runtime analysis of performance. We study a base-line, unbiased modularization that makes use of a complete module set (CMS), with special focus on strings that grow logarithmically with the problem size. We learn that even unbiased modularizations can have profound effects on problem performance. Our (1+1) CMS-EA optimizes a generalized OneMax problem in Ω(n2) time, provably worse than a (1+1) EA. More generally, our (1+1) CMS-EA optimizes a particular class of concatenated functions in O(2lm k n) time, where lm is the length of module strings and k is the number of module positions, when the modularization is aligned with the problem separability. We compare our results to known results for traditional EAs, and develop new intuition about modular encapsulation. We observe that search in the CMS-EA is essentially conducted at two levels (intra- and extra-module) and use this observation to construct a module trap, requiring super-polynomial time for our CMS-EA and O(n ln n) for the analogous EA.

This paper presents a communication-less multi-agent task allocation procedure that allows agents to use past experience to make non-greedy decisions about task assignments. Experimental results are given for problems where agents have varying capabilities, tasks have varying difficulties, and agents are ignorant of what tasks they will see in the future. These types of problems are difficult because the choice an agent makes in the present will affect the decisions it can make in the future. Current task-allocation procedures, especially the market-based ones, tend to side-step the issue by ignoring the future and assigning tasks to agents in a greedy way so that short-term goals are met. It is shown here that these short-sighted allocation procedures work well in situations where the ratio of task length to team size is small, but their performance decreases as this ratio increases. The adaptive method presented here is shown to perform well in a wide range of task-allocation problems, and because it requires no explicit communication, its computational costs are independent of team size.

Modularity is thought to improve the evolvability of biological systems [18, 22]. Recent studies in the field of evolutionary computation show that the use of modularity improves performance and scalability of evolutionary algorithms for certain applications. [5, 12, 15, 16, 17]. The effects of introducing modularity to evolutionary search, however, are not well understood. This paper focuses on analyzing the effects of modularity on evolutionary computation. In particular, we analyze the effects of modular representations on the search space bias.

This paper focuses on the evolution of ARTMAP architectures, using genetic algorithms, with the objective of improving generalization performance and alleviating the ART category proliferation problem. We refer to the resulting architectures as GFAM, GEAM, and GGAM. We demonstrate through extensive experimentation that evolved ARTMAP architectures exhibit good generalization and are of small size, while consuming reasonable computational effort to produce an optimal or a sub-optimal network. Furthermore, we compare the performance of GFAM, GEAM and GGAM with other competitive ARTMAP architectures that have appeared in the literature and addressed the category proliferation problem in ART. This comparison indicates that GFAM, GEAM and GGAM have superior performance (generalize better, are of smaller size, and require less computations) compared with other competitive ARTMAP architectures.

Fuzzy ARTMAP (FAM) is currently considered to be one of the premier neural network architectures in solving classification problems. One of the limitations of fuzzy ARTMAP that has been extensively reported in the literature is the category proliferation problem. That is fuzzy ARTMAP has the tendency of increasing its network size, as it is confronted with more and more data, especially if the data is of noisy and/or overlapping nature. To remedy this problem a number of researchers have designed modifications to the training phase of fuzzy ARTMAP that had the beneficial effect of reducing this phenomenon. In this paper we propose a new approach to handle the category proliferation problem in fuzzy ARTMAP by evolving trained FAM architectures. We refer to the resulting FAM architectures as GFAM We demonstrate through extensive experimentation that an evolved FAM (GFAM) exhibits good generalization, small size, and produces an optimal or a good sub-optimal network with a reasonable computational effort. Furthermore, comparisons of the GFAM with other approaches, proposed in the literature, that address the FAM category proliferation problem, illustrate that the GFAM has a number of advantages (i.e. produces smaller or equal size architectures, of better or as good generalization, with reduced computational complexity).

The execution of a complex task in any environment requires planning. Planning is the process of constructing an activity graph given by the current state of the system, a goal state, and a set of activities. If we wish to execute a complex computing task in a heterogeneous computing environment with autonomous resource providers, we should be able to adapt to changes in the environment. A possible solution is to construct a family of activity graphs beforehand and investigate the means of switching from one member of the family to another when the execution of one activity graph fails. In this paper, we study the conditions when plan switching is feasible. Then we introduce an approach for plan switching and report the simulation results of this approach

Data-parallel applications often use a mesh as the communication structure for their local-communication operations. To obtain the best possible performance, when executing on heterogeneous networks, it is important to match this mesh with the topology of the underlying network. This paper presents and compares strategies for matching meshes with heterogeneous networks. Our experiments have shown that some of the strategies presented do decrease the time spent in local communication and improve the performance of data-parallel applications

A key property for the effectiveness of stochastic search techniques, including evolutionary algorithms, is the existence of a positive correlation between the form and the quality of candidate solutions. In this paper, we show that when the ordering of genomic symbols in a genetic algorithm is completely independent of the fitness function and therefore free to evolve along the candidate solutions it encodes, the resulting genomes self-organize into self-similar structures that favor this key stochastic search property.

In this work we provide empirical evidence that shows how a variable-length genetic algorithm (GA) can naturally evolve shorter average size populations. This reduction in chromosome length appears to occur in finite population GAs when 1) selection is absent from the GA (random) or 2) when selection focuses on some other property not influenced by the length of individuals within a population.

This paper examines the efficacy of genetic algorithms (GAs) in combining input from multiple users to control a single interactive system, such as an educational exhibit at a museum. Specifically, the idea of collaborative interactive evolution (that is, interactive evolution with input from multiple users) is introduced for this purpose. Two fitness functions are proposed to guide the collaborative interactive evolution, as well as two non-GA methods for combining user input. The usefulness and success of each of these methods is examined, and the GA is shown to be a viable means for combining user input for the control of a single interactive system.

Conservation of functionally identical copies of the same gene throughout the generations is not an easy task. In this study, based on the biological evidence that suggests the existence of the developmental error as one of the ways to preserve redundancy, our goal is to investigate the impact of developmental error on a simple problem using a genetic algorithm(GA). Developmental errors exist during the biological development of an individual. The biological models with developmental errors have demonstrated that it is possible to maintain redundant copies in the existence of proper mutation rates and developmental error rates. Our preliminary results with a simple problem demonstrates that using developmental error helps preserving the redundant copies and maintaining a better solution quality for the redundant copies of a gene. Besides the developmental error, we propose a new error type that comes into play during the fitness assignment and enhances the quality of the solutions when used together with developmental error.

We have developed a genetic algorithm (GA) approach to the problem of task scheduling for multiprocessor systems. Our approach requires minimal problem specific information and no problem specific operators or repair mechanisms. Key features of our system include a flexible, adaptive problem representation and an incremental fitness function. Comparison with traditional scheduling methods indicates that the GA is competitive in terms of solution quality if it has sufficient resources to perform its search. Studies in a nonstationary environment show the GA is able to automatically adapt to changing targets.

Effective automatic control of Advanced Life Support Systems (ALSS) is a crucial component of space exploration. An ALSS is a coupled dynamical system which can be extremely sensitive and difficult to predict. As a result, such systems can be difficult to control using deliberative and deterministic methods. We investigate the performance of two machine learning algorithms, a genetic algorithm (GA) and a stochastic hill-climber (SH), on the problem of learning how to control an ALSS, and compare the impact of two different types of problem representations on the performance of both algorithms. We perform experiments on three ALSS optimization problems using five strategies with multiple variations of a proportional representation for a total of 120 experiments. Results indicate that although a proportional representation can effectively boost GA performance, it does not necessarily have the same effect on other algorithms such as SH. Results also support previous conclusions that multivector control strategies are an effective method for control of coupled dynamical systems.

We compare the traditional, fixed genetic algorithm (GA) representation scheme with a floating representation scheme and examine the differences in building block dynamics and how these differences affect a GA's ability to balance exploration and exploitation of building blocks. This study examines both the overall performance of a GA and the detailed events that contribute to overall behavior. Results indicate that the floating representation allows a GA to maintain a higher-level of construction which results in a more diverse population from which to build solutions.

The non-linear complexity of evolutionary algorithms (EAs) make them a challenge to understand. The difficulty in performing detailed analyses of an EA is in sorting through the large amount of of data that can be generated in a single run. This paper describes a visualization tool that facilitates navigation through the details of an EA run. The visualization tool organizes and displays EA data at various levels of detail and allows for easy transitions between related pieces of data.

The majority of current genetic algorithms (GAs), while inspired by natural evolutionary systems, are seldom viewed as biologically plausible models. This is not a criticism of GAs, but rather a reflection of choices made regarding the level of abstraction at which biological mechanisms are modeled, and a reflection of the more engineering-oriented goals of the evolutionary computation community. Understanding better and reducing this gap between GAs and genetics has been a central issue in an interdisciplinary project whose goal is to build GA-based computational models of viral evolution. The result is a system called Virtual Virus (VIV). VIV incorporates a number of more biologically plausible mechanisms, including a more flexible genotype-to-phenotype mapping. In VIV the genes are independent of position, and genomes can vary in length and may contain noncoding regions, as well as duplicative or competing genes. Initial computational studies with VIV have already revealed several emergent phenomena of both biological and computational interest. In the absence of any penalty based on genome length, VIV develops individuals with long genomes and also performs more poorly (from a problem-solving viewpoint) than when a length penalty is used. With a fixed linear length penalty, genome length tends to increase dramatically in the early phases of evolution and then decrease to a level based on the mutation rate. The plateau genome length (i.e., the average length of individuals in the final population) generally increases in response to an increase in the base mutation rate. When VIV converges, there tend to be many copies of good alternative genes within the individuals. We observed many instances of switching between active and inactive genes during the entire evolutionary process. These observations support the conclusion that noncoding regions serve as scratch space in which VIV can explore alternative gene values. These results represent a positive step in understanding how GAs might exploit more of the power and flexibility of biological evolution while simultaneously providing better tools for understanding evolving biological systems.

This article compares the traditional, fixed problem representation style of a genetic algorithm (GA) with a new floating representation in which the building blocks of a problem are not fixed at specific locations on the individuals of the population. In addition, the effects of noncoding segments on both of these representations is studied. Noncoding segments are a computational model of noncoding deoxyribonucleic acid, and floating building blocks mimic the location independence of genes. The fact that these structures are prevalent in natural genetic systems suggests that they may provide some advantages to the evolutionary process. Our results show that there is a significant difference in how GAs solve a problem in the fixed and floating representations. Genetic algorithms are able to maintain a more diverse population with the floating representation. The combination of noncoding segments and floating building blocks appears to encourage a GA to take advantage of its parallel search and recombination abilities.

The genetic algorithm (GA) is a problem-solving method that is modeled after the process of natural selection. We are interested in studying a specific aspect of the GA: the effect of noncoding segments on GA performance. Noncoding segments are segments of bits in an individual that provide no contribution, positive or negative, to the fitness of that individual. Previous research on noncoding segments suggests that including these structures in the GA may improve GA performance. Understanding when and why this improvement occurs will help us to use the GA to its full potential. In this article we discuss our hypotheses on noncoding segments and describe the results of our experiments. The experiments may be separated into two categories: testing our program on problems from previous related studies, and testing new hypotheses on the effect of noncoding segments.

We study a specific aspect of the genetic algorithm (GA): the effect of non-coding segments on GA performance. Non-coding segments are segments of bits in an individual that provide no contribution, positive or negative, to the fitness of that individual. Previous research on non-coding segments suggests that including these structures in the GA population may improve GA performance. As a first step in our research, we tested our program on some of the same problems as the previous studies. This paper compares our results with the previous results and discusses the significance of the similarities and differences.>