Obtaining high certainty in predictive models is crucial for making informed and trustworthy decisions in many scientific and engineering domains. However, extensive experimentation required for model accuracy can be both costly and time-consuming. This paper presents an adaptive sampling approach designed to reduce epistemic uncertainty in predictive models. Our primary contribution is the development of a metric that estimates potential epistemic uncertainty leveraging prediction interval-generation neural networks. This estimation relies on the distance between the predicted upper and lower bounds and the observed data at the tested positions and their neighboring points. Our second contribution is the proposal of a batch sampling strategy based on Gaussian processes (GPs). A GP is used as a surrogate model of the networks trained at each iteration of the adaptive sampling process. Using this GP, we design an acquisition function that selects a combination of sampling locations to maximize the reduction of epistemic uncertainty across the domain. We test our approach on three unidimensional synthetic problems and a multi-dimensional dataset based on an agricultural field for selecting experimental fertilizer rates. The results demonstrate that our method consistently converges faster to minimum epistemic uncertainty levels compared to Normalizing Flows Ensembles, MC-Dropout, and simple GPs.

Ant Colony Optimization (ACO) has served as a widely-utilized metaheuristic algorithm for decades for solving combinatorial optimization problems. Since its initial construction, ACO has seen a wide variety of modifications and connections to Reinforcement Learning (RL). Substantial parallels can be seen as early as 1995 with Ant-Q's relationship with Q-learning, through 2022 with ADACO's connection with Policy Gradient. In this work, we describe ACO, more specifically the Stochastic Gradient Descent ACO algorithm (ACOSGD), explicitly as an off-policy Policy Gradient (PG) method. We also incorporate experience replay into several ACO algorithm variants, including AS, MaxMin-ACO, ACOSGD, ADACO, and our two policy gradient-based versions: PGACO and PPOACO, drawing the connection to elitist ACO strategies. We show that our implementation of PG in ACO with experience replay and a baselined reward update strategy applied to eight TSP problems of varying sizes performs competitively with both fundamental ACO and SGD-based ACO versions. We also show that the replay buffer seems to unilaterally improve the performance of ACO algorithms through an ablation study.

Recent initiatives in United States university systems have been focusing on the problems associated with first-year students dropping out in high numbers. At Montana State University, a pilot program is underway to develop strategies for improving undergraduate student retention and reducing the time to graduation. For the pilot program, students found to be socioeconomically and academically disadvantaged are targeted to participate in the strategies designed to mitigate their disadvantaged state. Through the university’s "Persistence to Degree" initiative, these strategies have included forming student cohorts that take first-year classes together, thus promoting a sense of purpose and belonging. This paper presents one such strategy whereby the cohorts are formed based on mutual interests, and common first-year "core" courses (i.e., general education courses) are recommended for the members of the cohorts to take together. The approach involves employing a BERT-based topic model to generate a social network, from which communities are extracted. These communities are constructed based on student interests, expressed through a set of "minute essays," and a hybrid content-based/collaborative filtering method is employed to pair courses taken by past students with similar interests. Given that this project is still in the pilot stage, this paper focuses on the methodology and underlying ethical issues, rather than specific results since collecting results on the method’s effectiveness would require more of a longitudinal design. Even so, initial results show promise in the proposed methodology.

Accurate uncertainty quantification is necessary to enhance the reliability of deep learning (DL) models in real-world applications. In the case of regression tasks, prediction intervals (PIs) should be provided along with the deterministic predictions of DL models. Such PIs are useful or "high-quality (HQ)" as long as they are sufficiently narrow and capture most of the probability density. In this article, we present a method to learn PIs for regression-based neural networks (NNs) automatically in addition to the conventional target predictions. In particular, we train two companion NNs: one that uses one output, the target estimate, and another that uses two outputs, the upper and lower bounds of the corresponding PI. Our main contribution is the design of a novel loss function for the PI-generation network that takes into account the output of the target-estimation network and has two optimization objectives: minimizing the mean PI width and ensuring the PI integrity using constraints that maximize the PI probability coverage implicitly. Furthermore, we introduce a self-adaptive coefficient that balances both objectives within the loss function, which alleviates the task of fine-tuning. Experiments using a synthetic dataset, eight benchmark datasets, and a real-world crop yield prediction dataset showed that our method was able to maintain a nominal probability coverage and produce significantly narrower PIs without detriment to its target estimation accuracy when compared to those PIs generated by three state-of-the-art neural-network-based methods. In other words, our method was shown to produce higher quality PIs.

The Linear Model U-Tree (LMUT) has been used to increase the interpretability of Deep Reinforcement Learning (DRL) agents by mimicking behavior in terms of Q-value predictions and gameplay. In this paper, we consider two extensions to LMUT. First, we evaluate the impact of prepruning and bottomup postpruning on LMUT and find that while prepruning has a mixed to negligible impact on performance, postpruning brings its Q-value predictions closer in line with the DRL agent, increasing the effectiveness of its influence on DRL interpretability. Second, we find evidence that LMUT gameplay typically more closely matches that of the DRL agent it learns to mimic when the DRL agent policy is more robust to noise, even after controlling for the performance of the DRL agent on the underlying task. This indicates that LMUT efficacy is driven in part by the robustness of the DRL policy.

Community detection plays a pivotal role in social network analysis by partitioning networks into cohesive groups of vertices with dense intra-group connections and sparse inter-group connections. In this paper, we utilized a scholarly social network based on researchers' topic similarity derived from their publication metadata to identify interdisciplinary research communities. As topics often form a hierarchy, we hypothesize that the constructed scholarly network will exhibit hierarchical community structures. Therefore, we explore the efficacy of two prominent community detection algorithms, Louvain and Spectral clustering, known for their capacity to detect hierarchical community structures within networks. While both algorithms demonstrate this capability, the original Louvain algorithm is susceptible to the resolution limit problem due to its reliance on the modularity measure. To address this limitation, we propose the nested hierarchical Louvain algorithm, which iteratively partitions the network based on previously identified subgraphs, and we find that the bias towards large communities is mitigated. To evaluate the hierarchy produced by each of the algorithms, we employ the Cophenetic Correlation Coefficient (CPCC), a metric commonly used in hierarchical clustering evaluations but less frequently utilized in hierarchical community analysis. We argue that CPCC can be a useful measure to identify the presence of implicit hierarchical community structure in social networks when it is not explicitly available from domain knowledge while also further mitigating the inherent bias present in using modularity as a metric. Experimental results, conducted on both synthetic networks and the scholarly social network, demonstrate that the nested hierarchical Louvain algorithm, as well as Spectral Clustering, successfully identifies more finely structured hierarchical communities, offering greater depth in the dendrogram compared to the basic Louvain algorithm.

In Precision Agriculture, the utilization of management zones (MZs) that take into account within-field variability facilitates effective fertilizer management. This approach enables the optimization of nitrogen (N) rates to maximize crop yield production and enhance agronomic use efficiency. However, existing works often neglect the consideration of responsivity to fertilizer as a factor influencing MZ determination. In response to this gap, we present a MZ clustering method based on fertilizer responsivity. We build upon the statement that the responsivity of a given site to the fertilizer rate is described by the shape of its corresponding N fertilizer-yield response (N-response) curve. Thus, we generate N-response curves for all sites within the field using a convolutional neural network (CNN). The shape of the approximated N-response curves is then characterized using functional principal component analysis. Subsequently, a counter-factual explanation (CFE) method is applied to discern the impact of various variables on MZ membership. The genetic algorithm-based CFE solves a multi-objective optimization problem and aims to identify the minimum combination of features needed to alter a site’s cluster assignment. Results from two yield prediction datasets indicate that the features with the greatest influence on MZ membership are associated with terrain characteristics that either facilitate or impede fertilizer runoff, such as terrain slope or topographic aspect.

Recently, there has been a focus on using transfer learning to reduce the sample complexity in reinforcement learning. One component that enables transfer is an intertask mapping that relates a pair of tasks. Automatic methods attempt to learn task relationships either by evaluating all possible mappings in a brute force manner, or by using techniques such as neural networks to represent the mapping. However, brute force methods do not scale well in problems since there is an exponential number of possible mappings, and automatic methods that use complex representations generate mappings that are not always interpretable. In this paper, we describe a population-based algorithm that generates intertask mappings in a tractable amount of time. The idea is to use an explicit representation of an intertask mapping, and to combine an evolutionary algorithm with an offline evaluation scheme to search for the optimal mapping. Experiments on two transfer learning problems show that our approach is capable of finding highly-fit mappings and searching a space that is infeasible for a brute force approach. Furthermore, agents that learn using the mappings found by our approach are able to reach a performance target faster than agents that learn without transfer.

Uncertainty of the environment limits the circumstances with which any optimization problem can provide meaningful information. Multiple optimizers can combat this problem by communicating different information through cooperative coevolution. In reinforcement learning (RL), uncertainty can be reduced by applying learned policies collaboratively with another agent. Here, we propose policy Co-training with Factored Evolutionary Algorithms (CoFEA) to evolve an optimal policy for such scenarios. We hypothesize that self-paced co-training can allow factored particle swarms with imperfect knowledge to consolidate knowledge from each of their imperfect policies in order to approximate a single optimal policy. Additionally, we show how the performance of co-training swarms of RL agents can be maximized through the specific use of Expected SARSA as the policy learner. We evaluate CoFEA against comparable RL algorithms and attempt to establish limits for which our procedure does and does not provide benefit. Our results indicate that Particle Swarm Optimization (PSO) is effective in training multiple agents under uncertainty and that FEA reduces swarm and policy updates. This paper contributes to the field of cooperative co-evolutionary algorithms by proposing a method by which factored evolutionary techniques can significantly improve how multiple RL agents collaborate under extreme uncertainty to solve complex tasks faster than a single agent can under identical conditions.

Human Activity Recognition (HAR) is a difficult machine learning problem, even for state-of-the-art deep learning models, due to HAR data's within-domain and cross-domain heterogeneity. Our research addresses the challenge of closed-set domain adaptation in heterogeneous, parameter-based, and transductive transfer learning on HAR datasets. We use a Bidirectional Long Short Term Memory (BLSTM)-based model that, in addition to training for classification accuracy using only labeled data from the source domain, also jointly trains on source and unlabeled target datasets to reduce the discrepancy between source and target domains using cross-domain similarity as an additional loss function. Our work contributes to existing research in the area of domain adaptation for HAR by evaluating the performance of the following cross-domain similarity metrics as loss functions in improving model classification accuracy: 1) Maximum Mean Discrepancy (MMD), which uses feature means to measure similarity between two domains; 2) Kernel Canonical Correlation Analysis (KCCA), which utilizes canonical correlations for similarity determination; and 3) Cosine Similarity, a metric that uses the cosine of the angle between two vectors as similarity measure. Our results demonstrate that MMD as a cross-domain similarity metric not only outperforms KCCA and Cosine Similarity in domain adaption, but also results in mean F1 score improvement of 45% over results where a model is trained solely on the target dataset.

Response curves exhibit the magnitude of the response of a sensitive system to a varying stimulus. However, response of such systems may be sensitive to multiple stimuli (i.e., input features) that are not necessarily independent. As a consequence, the shape of response curves generated for a selected input feature (referred to as “active feature”) might depend on the values of the other input features (referred to as “passive features”). In this work we consider the case of systems whose response is approximated using regression neural networks. We propose to use counterfactual explanations (CFEs) for the identification of the features with the highest relevance on the shape of response curves generated by neural network black boxes. CFEs are generated by a genetic algorithm-based approach that solves a multi-objective optimization problem. In particular, given a response curve generated for an active feature, a CFE finds the minimum combination of passive features that need to be modified to alter the shape of the response curve. We tested our method on a synthetic dataset with 1-D inputs and two crop yield prediction datasets with 2-D inputs. The relevance ranking of features and feature combinations obtained on the synthetic dataset coincided with the analysis of the equation that was used to generate the problem. Results obtained on the yield prediction datasets revealed that the impact on fertilizer responsivity of passive features depends on the terrain characteristics of each field.

We propose a factored evolutionary framework for multi-objective optimization that can incorporate any multi-objective population based algorithm. Our framework, which is based on Factored Evolutionary Algorithms, uses overlapping subpopulations to increase exploration of the objective space; however, it also allows for the creation of distinct subpopulations as in co-operative co-evolutionary algorithms (CCEA). We apply the framework with the Non-Dominated Sorting Genetic Algorithm-II (NSGA-II), resulting in Factored NSGA-II. We compare NSGA-II, CC-NSGA-II, and F-NSGA-II on two different versions of the multi-objective knapsack problem. The first is the classic binary multi-knapsack implementation introduced by Zitzler and Thiele, where the number of objectives equals the number of knapsacks. The second uses a single knapsack where, aside from maximizing profit and minimizing weight, an additional objective tries to minimize the difference in weight of the items in the knapsack, creating a balanced knapsack. We further extend this version to minimize volume and balance the volume. The proposed 3-to-5 objective balanced single knapsack problem poses a difficult problem for multi-objective algorithms. Our results indicate that the non-dominated solutions found by F-NSGA-II tend to cover more of the Pareto front and have a larger hypervolume.

Discovery and partitioning of complex graphs such as social networks into communities can be useful for analyzing behavior of individuals in the network. Spectral clustering is one useful tool for performing this clustering, but it suffers from scalability and generalizability to new data. In this paper we introduce an approximate orthogonal spectral autoencoder. We apply this model to a political campaign contribution social network to show its effectiveness on out-of-sample embedding for clustering and classification. The resulting embedding is used with hierarchical fuzzy spectral clustering to show embedding generalization for a behavioral prediction problem for nodes in the social network.

The large size of hyperspectral images limits the applicable uses and necessitates effective compression methods. While there has been great success in a combined spectral and spatial compression approach, the reconstruction error rate for lossy compression methods with higher compression rates is still relatively large. Inspired by recent successes in spectral based deep learning for classification, we propose a spectral based Long Short Term Memory Autoencoder (LSTM-AE) to compress the spectral dimension alone. The obtained results show that not only can LSTM-AE achieve similar compression rates to existing methods, but also a large reduction in reconstruction error. We have also demonstrated the robustness of the approach in being able to generalize a single model for use in multiple scenes without being retrained. The model was also demonstrated to be successful in compressing unseen images at higher rates than existing methods that have trained on those images.

Particle swarm optimization (PSO) has become a popular algorithm for performing global numerical optimization; however, it is known that the topology of PSO has a large influence on its performance. Topologies with high connectivity can have fast convergence, but they are also susceptible to convergence to local minima. Topologies with low connectivity may avoid converging to local minima and achieve high quality solutions, but they tend to have slow convergence. In this paper, we propose a novel PSO topology based on a single-elimination tournament. In the proposed tournament topology, particles move up a tree structure through a fitness-based tournament. PSO updates then propagate information about the global best position from the top of the tree to the bottom. Experimental results on eleven benchmark functions show that the proposed topology can achieve both the high quality solutions of low-connectivity topologies and the fast convergence of high-connectivity topologies.

In recent years, Hyperspectral Imaging (HSI) has become a powerful source for reliable data in applications such as remote sensing, agriculture, and biomedicine. However, hyperspectral images are highly data-dense and often benefit from methods to reduce the number of spectral bands while retaining the most useful information for a specific application. We propose a novel band selection method to select a reduced set of wavelengths, obtained from an HSI system in the context of image classification. Our approach consists of two main steps: the first utilizes a filter-based approach to find relevant spectral bands based on a collinearity analysis between a band and its neighbors. This analysis helps to remove redundant bands and dramatically reduces the search space. The second step applies a wrapper-based approach to select bands from the reduced set based on their information entropy values, and trains a compact Convolutional Neural Network (CNN) to evaluate the performance of the current selection. We present classification results obtained from our method and compare them to other feature selection methods on two hyperspectral image datasets. Additionally, we use the original hyperspectral data cube to simulate the process of using actual filters in a multispectral imager. We show that our method produces more suitable results for a multispectral sensor design.

Some data exhibit natural divisions where the application of a single neural network leaves some accuracy on the table, thereby making a multi-network approach more appropriate. We develop an approach to preserving knowledge encoded in the hidden layer of several ANN's and assemble that knowledge in new, composite networks based on spatial clustering that more effectively make predictions over subdivisions of the entire dataspace. We show that this method has an accuracy advantage over the single-network approach.

Quantifying uncertainty is critically important to many applications of predictive modeling. In this paper we apply a recently developed method that uses U-statistics as a basis for estimating uncertainty in ensemble regressors to the case of neural network ensembles. U-statistics generalize the notion of a sample mean and provide distributional properties to estimates obtained by ensembles of estimators. With this method, we train neural networks on subsamples of the data and use the resulting ensemble to estimate the variance of the point estimates from the ensemble. We demonstrate that neural networks predicting a regression function exhibit the required theoretical properties for use in this ensemble method, and we then perform a coverage probability study of three simulated data sets to show that the empirical coverage probabilities match the theoretical values.

In this paper, we seek to automate the evaluation of explanations of image classification decisions made by complex convolutional neural networks (CNN). Explanation frameworks like Local Interpretable Model-agnostic Explanations (LIME) treat complex machine learning models, such as deep neural networks, as black boxes and generate human-interpretable explanations of their decisions using linear proxy models. We propose a pair of experiments to quantitatively evaluate the quality of generated explanations by measuring their sufficiency and salience. To test if a generated explanation contains sufficient information for classification, we test the ability of a trained CNN to classify that explanation properly. We test explanations for salience by training two new CNNs, one using raw image data and the other using explanations as training data, and comparing their classification precision and recall on a common set of test data. We use our new evaluation framework to test our hypothesis that LIME is able to generate explanations that are both sufficient and salient. Our results show that the generated explanations have the potential to be sufficient and salient, provided that the complexity of the explanations is enough to describe the underlying classes.

Variable Rate Application (VRA) is a popular technique in Precision Agriculture used to decrease the amount of fertilizer applied to a specific field while increasing profitability, effectively also reducing environmental impact. VRA tries to determine the rate of fertilizer to apply to different parts of a field based on a variety of factors, such as precipitation, elevation, and previous years' yield. To determine the appropriate variable nitrogen application rate for a field, experiments have to be conducted that provide data on how certain parts of the field react to specific nitrogen rates. In this research, a VRA of nitrogen is applied to fields of winter wheat in Montana where these experiments require the creation of a prescription map, which creates a grid of the field. The goal of the experiments is to vary nitrogen rate application, to determine how these nitrogen rates affect yield and protein production. However, when creating these prescription maps large jumps between consecutive cells' nitrogen rates often occur, putting strain on the farming equipment. To reduce the number of jumps while maintaining even distribution of nitrogen rates across different yield and protein bins, a Genetic Algorithm (GA) is used for optimization. The GA uses a multi-objective fitness function aiming to minimize jumps and maintain stratification. The results show that the GA is effective in meeting these goals for the fields studied.

Optical sensing has the potential to be an important tool in the automated monitoring of food quality. Specifically, hyperspectral imaging has enjoyed success in a variety of tasks ranging from plant species classification to ripeness evaluation in produce. Although effective, hyperspectral imaging is prohibitively expensive to deploy at scale in a retail setting. With this in mind, we develop a method to assist in designing a low-cost multispectral imager for produce monitoring by using a genetic algorithm (GA) that simultaneously selects a subset of informative wavelengths and identifies effective filter bandwidths for such an imager. Instead of selecting the single fittest member of the final population as our solution, we fit a univariate Gaussian mixture model to the histogram of the overall GA population, selecting the wavelengths associated with the peaks of the distributions as our solution. By evaluating the entire population, rather than a single solution, we are also able to specify filter bandwidths by calculating the standard deviations of the Gaussian distributions and computing the full-width at half-maximum values. In our experiments, we find that this novel histogram-based method for feature selection is effective when compared to both the standard GA and partial least squares discriminant analysis.

An important aspect of social networks is the discovery and partitioning of the complex graphs into dense sub-networks referred to as communities. The goal of such partitioning is to find groups who have similar attributes or behaviors. In the realm of politics, it is possible to group individuals with similar political behavior by analyzing campaign finance records. In this paper we use fuzzy hierarchical spectral clustering to find communities with campaign finance networks. Multiple experiments were performed using varying edge weighting, number and type of communities, as well as analyzing multiple different years of voting data. The results show that using the full hierarchy of community assignments for legislators is highly predictive of voting behavior in the US House of Representatives and Senate.

Recent advances in machine learning have brought with them considerable attention in applying such methods to complex prediction problems. However, in extremely large dataspaces, a single neural network covering that space may not be effective, and generating large numbers of deep neural networks is not feasible. In this paper, we analyze deep networks trained from stacked autoencoders in a spatio-temporal application area to determine the extent to which knowledge can be transferred to similar regions. Our analysis applies methods from functional data analysis and spatial statistics to identify such correlation. We apply this work in the context of numerical weather prediction in analyzing large-scale data from Hurricane Sandy. Results of our analysis indicate high likelihood that spatial correlation can be exploited if it can be identified prior to training.

Adaptive Boosting, or AdaBoost, is an algorithm aimed at improving the performance of ensembles of weak learners by weighing the data itself as well as the learners. Two versions of AdaBoost-AdaBoost-R2 and AdaBoost RΔ-are applied in this project, as well as a third novel algorithm combining ideas of these two methods, to the problem of predicting crop yield and protein content in support of precision agriculture. All three algorithms use Feedforward Neural Networks (FFNN) trained with backpropagation as the weak model. Data from four different fields were gathered as a result of on-farm experiments of different nitrogen rate applications using randomly stratified trials based on previous years' yield and protein. The three AdaBoost algorithms are compared to a simple FFNN with a single hidden layer. The results confirm previous findings in different fields, where ensemble methods outperform single models. The results are improved by 3 to 10 units for yield prediction, and by a small percentage for protein prediction.

While a great deal of research has been directed towards developing neural network architectures for RGB images, there is a relative dearth of research directed towards developing neural network architectures specifically for multi-spectral and hyper-spectral imagery. We have adapted recent developments in small efficient convolutional neural networks (CNNs), to create a small CNN architecture capable of being trained from scratch to classify 10 band multi-spectral images, using much fewer parameters than popular deep architectures, such as the ResNet or DenseNet architectures. We show that this network provides higher classification accuracy and greater sample efficiency than the same network using RGB images. Further, using a Bayesian version of our CNN architecture we show that a network that is capable of working with multi-spectral imagery significantly reduces the uncertainty associated with class predictions compared to using RGB images.

Neural network pruning can be an effective method for creating more efficient networks without incurring a significant penalty in accuracy. It has been shown that the topology induced by pruning after training can be used to re-train a network from scratch on the same data set, with comparable or better performance. In the context of convolutional neural networks, we build on this work to show that not only can networks be pruned to 10% of their original parameters, but that these sparse networks can also be re-trained on similar data sets with only a slight reduction in accuracy. We use the Lottery Ticket Hypothesis as the basis for our pruning method and discuss how this method can be an alternative to transfer learning, with positive initial results. This paper lays the groundwork for a transfer learning method that reduces the original network to its essential connections and does not require freezing entire layers.

Particle Swarm Optimization is an effective stochastic optimization technique that simulates a swarm of particles that fly through a problem space. In the process of searching the problem space for a solution, the individual variables of a candidate solution will often take on inferior values characterized as “Two Steps Forward, One Step Back.” Several approaches to solving this problem have introduced varying notions of cooperation and competition. Instead we characterize the success of these multi-swarm techniques as reconciling conflicting information through a mechanism that makes successive candidates Pareto improvements. We use this analysis to construct a variation of PSO that applies this mechanism to gbest selection. Experiments show that this algorithm performs better than the standard gbest PSO algorithm.

Competition and cooperation are powerful metaphors that have informed improvements in multi-population algorithms such as the Cooperative Coevolutionary Genetic Algorithm, Cooperative Particle Swarm Optimization, and Factored Evolutionary Algorithms (FEA). However, we suggest a different perspective can give a finer grained understanding of how multi-population algorithms come together to avoid problems like hitchhiking and pseudo-minima. In this paper, we apply the concepts of information sharing and conflict resolution through Pareto improvements to analyze the distributed version of FEA (DFEA). As a result, we find the original DFEA failed to implement FEA with complete fidelity. We then revise DFEA and examine the differences between it and FEA and the new implications for relaxing consensus in the distributed algorithm.

Factored Evolutionary Algorithms (FEA) have been found to be an effective way to optimize single objective functions by partitioning the variables in the function into overlapping subpopulations, or factors. While there exist several works empirically evaluating FEA, there exists very little literature exploring FEA's theoretical properties. In this paper, we prove that the final solution returned by FEA will be the results of converging to a single point. Additionally, we show how the convergence of FEA to a single point in the search space could be to a suboptimal point in space. However, we demonstrate empirically that when using specific factor architectures, the probability of converging to these suboptimal points in space approaches zero. Finally, where hybrid versions Cooperative Coevolutionary Algorithms have been proposed as a means to escape these suboptimal points, we show how FEA is able to outperform its hybrid version.

Factored evolutionary algorithms (FEAs) are a new class of evolutionary search-based optimization algorithms that have successfully been applied to various problems, such as training neural networks and performing abductive inference in graphical models. An FEA is unique in that it factors the objective function by creating overlapping subpopulations that optimize over a subset of variables of the function. In this paper, we give a formal definition of FEA algorithms and present empirical results related to their performance. One consideration in using an FEA is determining the appropriate factor architecture, which determines the set of variables each factor will optimize. For this reason, we present the results of experiments comparing the performance of different factor architectures on several standard applications for evolutionary algorithms. Additionally, we show that FEA's performance is not restricted by the underlying optimization algorithm by creating FEA versions of hill climbing, particle swarm optimization, genetic algorithm, and differential evolution and comparing their performance to their single-population and cooperative coevolutionary counterparts.

This paper describes an extension to overlapping swarm intelligence for training artificial neural networks. Overlapping swarm intelligence is an application of particle swarm optimization that divides the network into paths from input to output, with each path represented by a swarm. Previous versions of this algorithm showed success on training networks on a variety of datasets but the method suffers from an explosion in fitness evaluations due to the number of paths that need to be evaluated. We propose an extension to overlapping swarm intelligence to use asynchronous updates and dynamic subsets of swarms for each generation, and demonstrate that this method performs as well as basic overlapping swarm intelligence in terms of mean squared error and classification accuracy with fewer fitness evaluations.

Factored Evolutionary Algorithms (FEA) have proven to be fast and efficient optimization methods, often outperforming established methods using single populations. One restriction to FEA is that it requires a central communication point between all of the factors, making FEA difficult to use in completely distributed settings. The Distributed Factored Evolutionary Algorithm (DFEA) relaxes this requirement on central communication by having neighboring factors communicate directly with one another. While DFEA has been effective at finding good solutions, there is often an increase in computational complexity due to the communication between factors. In previous work on DFEA, the authors required the algorithm reach full consensus between factors during communication. In this paper, we demonstrate that even without full consensus, the performance of DFEA was not statistically different on problems with low epistasis. Additionally, we found that there is a relationship between the convergence of consensus between factors and the convergence of fitness of DFEA.

Particle Swarm Optimization (PSO) has been shown to perform very well on a wide range of optimization problems. One of the drawbacks to PSO is that the base algorithm assumes continuous variables. In this paper, we present a version of PSO that is able to optimize over discrete variables. This new PSO algorithm, which we call Integer and Categorical PSO (ICPSO), incorporates ideas from Estimation of Distribution Algorithms (EDAs) in that particles represent probability distributions rather than solution values, and the PSO update modifies the probability distributions. In this paper, we describe our new algorithm and compare its performance against other discrete PSO algorithms. In our experiments, we demonstrate that our algorithm outperforms comparable methods on both discrete benchmark functions and NK landscapes, a mathematical framework that generates tunable fitness landscapes for evaluating EAs.

Deep learning has recently gained popularity in many machine learning applications, but a theoretical grounding for the strengths, weaknesses, and implicit biases of various deep learning methods is still a work in progress. Here, we analyze the role of spatial locality in Deep Belief Networks (DBN) and show that spatially local information is lost through diffusion as the network becomes deeper. We then analyze an approach we developed previously, based on partitioning of Restricted Boltzmann Machines (RBMs), to demonstrate that our method is capable of retaining spatially local information when training DBNs. Specifically, we find that spatially local features are completely lost in DBNs trained using the “standard” RBM method, but are largely preserved using our partitioned training method. In addition, reconstruction accuracy of the model is improved using our Partitioned-RBM training method.

We develop a Partitioned Restricted Boltzmann Machine (PRBM) for classification. We demonstrate that this method provides both speed and accuracy. Specifically, because it is partitioned into smaller RBMs, all available data can be used for training, and individual RBMs can be trained in parallel. Moreover, as the number of dimensions increases, the number of partitions can be increased to significantly reduce runtime computational resource requirements. All other recently developed methods using RBMs for classification suffer from some serious disadvantage under bounded computational resources; one is forced to either use a subsample of the whole data, run fewer iterations (early stop criterion), or both. Our Partitioned-RBM method provides an innovative scheme to overcome this shortcoming.

Bayesian networks are probabilistic graphical models that have proven to be able to handle uncertainty in many real-world applications. One key issue in learning Bayesian networks is parameter estimation, i.e., learning the local conditional distributions of each variable in the model. While parameter estimation can be performed efficiently when complete training data is available (i.e., when all variables have been observed), learning the local distributions becomes difficult when latent (hidden) variables are introduced. While Expectation Maximization (EM) is commonly used to perform parameter estimation in the context of latent variables, EM is a local optimization method that often converges to sub-optimal estimates. Although several authors have improved upon traditional EM, few have applied population based search techniques to parameter estimation, and most existing population-based approaches fail to exploit the conditional independence properties of the networks. We introduce two new methods for parameter estimation in Bayesian networks based on particle swarm optimization (PSO). The first is a single swarm PSO, while the second is a multi-swarm PSO algorithm. In the multi-swarm version, a swarm is assigned to the Markov blanket of each variable to be estimated, and competition is held between overlapping swarms. Results of comparing these new methods to several existing approaches indicate that the multi-swarm algorithm outperforms the competing approaches when compared using data generated from a variety of Bayesian networks.

Deep learning is a popular field that encompasses a range of multi-layer connectionist techniques. While these techniques have achieved great success on a number of difficult computer vision problems, the representation biases that allow this success have not been thoroughly explored. In this paper, we examine the hypothesis that one strength of many deep learning algorithms is their ability to exploit spatially local statistical information. We present a formal description of how data vectors can be partitioned into sub-vectors that preserve spatially local information. As a test case, we then use statistical models to examine how much of such structure exists in the MNIST dataset. Finally, we present experimental results from training RBMs using partitioned data, and demonstrate the advantages they have over non-partitioned RBMs. Through these results, we show how the performance advantage is reliant on spatially local structure, by demonstrating the performance impact of randomly permuting the input data to destroy local structure. Overall, our results support the hypothesis that a representation bias reliant upon spatially local statistical information can improve performance, so long as this bias is a good match for the data. We also suggest statistical tools for determining a priori whether a dataset is a good match for this bias or not.

Bayesian networks are powerful probabilistic models that have been applied to a variety of tasks. When applied to classification problems, Bayesian networks have shown competitive performance when compared to other state-of-the-art classifiers. However, structure learning of Bayesian networks has been shown to be NP-Hard. In this paper, we propose a novel approximation algorithm for learning Bayesian network classifiers based on Overlapping Swarm Intelligence. In our approach a swarm is associated with each attribute in the data. Each swarm learns the edges for its associated attribute node and swarms that learn conflicting structures compete for inclusion in the final network structure. Our results indicate that, in many cases, Overlapping Swarm Intelligence significantly outperforms competing approaches, including traditional particle swarm optimization.

The use of swarm robotics in search tasks is an active area of research. A variety of algorithms have been developed that effectively direct robots toward a desired target by leveraging their collaborative sensing capabilities. Unfortunately, these algorithms often neglect the task of communicating possible task solutions outside of the swarm. Many scenarios require a monitoring station that must receive updates from robots within the swarm. This task is trivial in constrained locations, but becomes difficult as the search area increases and communication between nodes is not always possible. A second shortcoming of existing algorithms is the inability to find and track mobile targets. We propose an extension to the distributed Particle Swarm Optimization algorithm that is both communication-aware and capable of tracking mobile targets within a search space. Simulated experiments show that our algorithm returns more accurate solutions to a monitoring station than existing algorithms, especially in scenarios, where the target value or location changes over time.

Abductive inference in Bayesian networks, is the problem of finding the most likely joint assignment to all non-evidence variables in the network. Such an assignment is called the most probable explanation (MPE). A novel swarm-based algorithm is proposed that finds the k-MPE of a Bayesian network. Our approach is an overlapping swarm intelligence algorithm in which a particle swarm is assigned to each node in the network. Each swarm searches for value assignments for its node's Markov blanket. Swarms that have overlapping value assignments compete to determine which assignment will be used in the final solution. In this paper we compare our algorithm to several other local search algorithms and show that our approach outperforms the competing methods in its ability to find the k-MPE.

Lexical abstraction hierarchies can be leveraged to provide semantic information that characterizes features of text corpora as a whole. This information may be used to determine the classification utility of the dimensions that describe a dataset. This paper presents a new method for preparing a dataset for probabilistic classification by determining, a priori, the utility of a very small subset of taxonomically-related dimensions via a Discriminative Multinomial Naive Bayes process. We show that this method yields significant improvements over both Discriminative Multinomial Naive Bayes and Bayesian network classifiers alone.

Deep structure learning is a promising new area of work in the field of machine learning. Previous work in this area has shown impressive performance, but all of it has used connectionist models. We hope to demonstrate that the utility of deep architectures is not restricted to connectionist models. Our approach is to use simple, non-connectionist dimensionality reduction techniques in conjunction with a deep architecture to examine more precisely the impact of the deep architecture itself. To do this, we use standard PCA as a baseline and compare it with a deep architecture using PCA. We perform several image classification experiments using the features generated by the two techniques, and we conclude that the deep architecture leads to improved classification performance, supporting the deep structure hypothesis.

The process of automatically tracking people within video sequences is currently receiving a great deal of interest within the computer vision research community. In this paper we contrast the performance of the popular Mean-Shift algorithmpsilas gradient descent based search strategy with a more advanced swarm intelligence technique. Towards this end, we propose the use of a Particle Swarm Optimization (PSO) algorithm to replace the gradient descent search, and also combine the swarm based search strategy with a Probabilistic Data Association Filter (PDAF) state estimator to perform the track association and maintenance stages. Performance is shown against a variety of data sets, ranging from easy to complex. The PSO-PDAF approach is seen to outperform both the Mean-Shift + Kalman filter and the single-measurement PSO + Kalman filter approach. However, PSOpsilas robustness to low contrast and occlusion comes at the cost of higher computational requirements.

The study of artificial life involves simulating biological or sociological processes with a computer. Combining artificial life with techniques from evolutionary computation frequently involves modeling the behavior or decision processes of artificial organisms within a society in such a way that genetic algorithms can be applied to modify these models and enhance behavior over time. Typically, endogenous fitness is used with co-evolution. We explore the use of an exogenous fitness function with genetic programming and co-evolution to develop individuals and species capable of competing in a hostile environment. To facilitate the study, we use a commercially available environment-AI Wars-to host the organisms and run the experiments. Results from our experiments, though preliminary, indicate the ability of co-evolution, genetic programming, and exogenous fitness to evolve fit individuals. The results also suggest the ability to assess the nature of the fitness landscape and the impact of various fitness factors on evolutionary performance.

We present a description of the AI-ESTATE (IEEE 1232.1) standard for representing and exchanging diagnostic models. These models are based on accepted approaches to performing system diagnostics in both commercial and military environments. Specifically, we discuss a standard approach to representing diagnostic fault trees and enhanced diagnostic inference models.

We describe a recently approved IEEE standard for exchanging diagnostic information and embedding diagnostic reasoners in any test environment. We describe the defined formats and services, an example application, and current industry acceptance.

Using nearest neighbor classification with fault dictionaries to resolve inexact signature matches in digital circuit diagnosis is inadequate. Nearest neighbor focuses on the possible diagnoses rather than on the tests. Our alternative-the information flow model-focuses on test information in the fault dictionary to provide more accurate diagnostics.

Testability is a measure of the potential to evaluate performance, determine operability, or identify faults within a system. Testability must be coupled with a strategy by which it is used to achieve field maintainability. It is often difficult to distinguish between shortcomings in testability and shortcomings in diagnostic strategy; however, adequate levels of both testability and diagnosis are required. An analysis of the Blackhawk helicopter illustrates the importance of matching the maintenance procedures to the maintenance architecture and the mission requirements.>

Improved interface standards and reduced design budgets dictate that commercial off-the-shelf (COTS) equipment be more readily integrated into system design. Often COTS is chosen for its functional capabilities and electronic compatibilities with little regard to testability and maintainability features. COTS equipment is often characterized by a lack of detailed information about the specific internal design of the equipment. Complex interactions across an array of subsystems may decrease the diagnosability of the system as a whole where deficits in information occur. In this paper, we describe an analysis approach for assessing system testability and providing system diagnostics that is amenable to including COTS equipment in the system under test. We illustrate the approach with the standard analysis of a system consisting of several subsystems with full information available.>

In diagnosing a failed system, a smart technician would choose tests to be performed based on the context of the situation. Currently, test program sets do not fault-. isolate within the context of a situation. Instead, testing follows a rigid, predetermined, fault-isolation sequence that is based on an embedded fault tree. Current test programs do not tolerate instrument failure and cannot redirect testing by incorporating new information. However, there is a new approach to automatic testing that emulates the best features of a trained technician yet, unlike the development of rule-based expert systems, does not require a trained technician to build the knowledge base. This new approach is model-based and has evolved over the last 10 years. This evolution has led to the development of several maintenance tools and an architecture for intelligent automatic test equipment (ATE). The architecture has been implemented for testing two cards from an AV-8B power supply.