Logical Table -to-Text (LT2T) generation aims to produce natural-language sentences that are logically faithful to structured tabular data.While recent Large Language Models (LLMs) show high performance on aggregate fidelity metrics, these scores provide only a coarse view of performance, obscuring specific logic-type reasoning failures and models' meta-logical awareness.We propose an operation-aware diagnostic framework that evaluates four core competencies: (1) Logical Form (LF) execution accuracy, (2) fidelity of LF-conditioned generation, (3) logic-type identification, and (4) LF-free generation.We apply this framework to a suite of frontier LLMs and perform fine-grained analysis across logic types such as aggregation, ordinal, and superlative reasoning.Our results show that LT2T fidelity assessment can be unstable; the choice of verifier and logic type can substantially alter conclusions and model rankings.Crucially, we identify a meta-logical gap: models often generate faithful statements while failing to identify the underlying operation.Logical Form: ( avg { all_rows ; score } ; 317.8) Task2's output LF text: "The average score of the leading rugby union test point scorers is almost 318 ."LLM model Task3's output Logic Type: Aggregation Reasoning: The core function is avg, an aggregation function. Task1's OutputFunction: average, Arguments: [all_rows, score], Input_rows: [0, 1, 2, 3, 4], Output_rows: [], Evidence_cells: [ {"row": 0, "col": 4, "value": "389"}, {"row": 1, "col": 4, "value": "335"}, {"row": 2, "col": 4, "value": "316"}, {"row": 3, "col": 4, "value": "279"}, {"row": 4, "col": 4, "value": "270"}], column: 4 Result: 317.8

In the field of machine learning, understanding how well deep neural networks (DNNs) are likely to learn certain features from training is essential for predicting model behaviors. We empirically study how a solution space’s feature-imbalance shapes learned distributions of mean-absolute error (MAE) and mean-bias error (MBE), and also consider the impact of loss functions, minor problem space perturbations, and localized quantile views within the solution space. Experiments were performed using supervised entity tracking problems designed to control for these variables and support broad testing. DNNs were trained using several common losses as baselines: the mean squared error (MSE), binary cross-entropy (BCE), and Huber. Additionally, we introduce and apply the novel "balanced type error" (BTE) loss, which computes a weighted average of errors relative to the feature-imbalance. MAE distributions remain stable across loss and problem space configurations, varying only with feature-imbalance and quantile view. MBE distributions, by contrast, are more unstable and additionally vary across loss and problem space configurations. Importantly, the BTE demonstrates independence between MAE and MBE distributions. Finally, although distribution trends are clear, noise in the empirical observations prevents clean generalizations.

The growth of machine learning demands interpretable models for critical applications, yet most high-performing models are "black-box" systems that obscure input-output relationships, while traditional rule-based algorithms like RuleFit suffer from a lack of predictive power and instability despite their simplicity. This motivated our development of Sparse Relaxed Regularized Regression Rule-Fit (SR4-Fit), a novel interpretable classification algorithm that addresses these limitations while maintaining superior classification performance. Using demographic characteristics of U.S. congressional districts from the Census Bureau’s American Community Survey, we demonstrate that SR4-Fit can predict House election party outcomes with unprecedented accuracy and interpretability. Our results show that while the majority party remains the strongest predictor, SR4-Fit has revealed intrinsic combinations of demographic factors that affect prediction outcomes that were unable to be interpreted in black-box algorithms such as random forests. The SR4-Fit algorithm surpasses both black-box models and existing interpretable rule-based algorithms such as RuleFit with respect to accuracy, simplicity, and robustness, generating stable and interpretable rule sets while maintaining superior predictive performance, thus addressing the traditional trade-off between model interpretability and predictive capability in electoral forecasting. To further validate SR4-Fit’s performance, we also apply it to six additional publicly available classification datasets, like the breast cancer, Ecoli, page blocks, Pima Indians, vehicle, and yeast datasets, and find similar results.

Previous work in Neural Loss Function Search (NLFS) has shown a lack of correlation between smaller surrogate functions and large convolutional neural networks with massive regularization. We expand upon this research by revealing another disparity that exists, correlation between different types of image augmentation techniques. We show that different loss functions can perform well on certain image augmentation techniques, while performing poorly on others. We exploit this disparity by performing an evolutionary search on five types of image augmentation techniques in the hopes of finding image augmentation specific loss functions. The best loss functions from each evolution were then taken and transferred to WideResNet-28-10 on CIFAR-10 and CIFAR-100 across each of the five image augmentation techniques. The best from that were then taken and evaluated by fine-tuning EfficientNetV2Small on the CARS, Oxford-Flowers, and Caltech datasets across each of the five image augmentation techniques. Multiple loss functions were found that outperformed cross-entropy across multiple experiments. In the end, we found a single loss function, which we called the inverse bessel logarithm loss, that was able to outperform cross-entropy across the majority of experiments.

A major contributor to the quality of a deep learning model is the selection of the optimizer. We propose a new dual-joint search space in the realm of neural optimizer search (NOS), along with an integrity check, to automate the process of finding deep learning optimizers. Our dual-joint search space simultaneously allows for the optimization of not only the update equation, but also internal decay functions and learning rate schedules for optimizers. We search the space using our proposed mutation-only, particle-based genetic algorithm able to be massively parallelized for our domain-specific problem. We evaluate our candidate optimizers on the CIFAR-10 dataset using a small ConvNet. To assess generalization, the final optimizers were then transferred to large-scale image classification on CIFAR-100 and TinyImageNet, while also being fine-tuned on Flowers102, Cars196, and Caltech101 using EfficientNetV2Small. We found multiple optimizers, learning rate schedules, and Adam variants that outperformed Adam, as well as other standard deep learning optimizers, across the image classification tasks.

The COVID-19 pandemic is a worldwide crisis with impacts that are both devastating and inequitable as effects often fall hardest on communities that are already suffering from economic, social, and political disparities. Interpretable machine learning (IML) offers the possibility for detailed understanding of this and similar disease outbreaks, allowing subject matter experts to explore the data more thoroughly and find patterns and connections that might otherwise remain hidden. As an active area of research in artificial intelligence, IML has great significance yet numerous technical challenges to overcome. In this paper, we focus on approximating epidemic curves using an interpretable artificial neural network. This is a first step toward a flexible and interpretable modeling framework that we plan to use to study impacts of various demographic, socioeconomic, and other factors on disease outbreaks. We tap into a substantial but little-known collection of IML studies in nonlinear function approximation from engineering mechanics, where domain knowledge including visually observable features of the data is systematically sorted and directly utilized in the initialization of sigmoidal neural networks leading to training success and good generalization. After an introductory review of existing work, we present a feasibility study on approximating a particular epidemic curve leading to a promising result.

Using a selection of machine learning methods and characteristic demographics of U.S. Congressional districts, we aim to determine if party outcome of a House election can be predicted from the demographics of the corresponding district. Utilizing demographics from the Census Bureau's annual American Community Survey and training models on these features we show that while it is possible to predict with high accuracy the party elected in a district, the strongest predictor remains the incumbent party. Further, we observe that certain demographic characteristics such as race and education have much stronger predictive power than age, income, and marital status. Nonetheless, even these weak predictors can enhance model accuracy when properly combined. We also consider model interpretability and informativeness, focusing on decision trees.

Reinforcement learning in continuous action spaces requires mechanisms that allow for exploration of infinite possible actions. One challenging issue in such systems is the amount of exploration appropriate during learning. This issue is complicated further in sharply changing dynamic environments. Reinforcement learning in artificial neural networks with multiparameter distributions can address all aspects of these issues. However, which equations are most appropriate for updating these parameters remains an open question. Here we consider possible equations derived from two sources: The classic equations proposed for REINFORCE and modern equations introduced for Stochastic Synapse Reinforcement Learning (SSRL), as well as combinations thereof and variations thereon. Using a set of multidimensional robot inverse kinematics problems, we find that novel combinations of these equations outperform either set of equations alone in terms of both learning rate and consistency.

Speciation is a critical process in both evolutionary biology and evolutionary computation. It helps us understand the development of life on Earth and can improve evolutionary algorithms' abilities to find global maxima. In recent years, there have been considerable efforts to understand this process and uncover new techniques to simulate it. While there are many approaches for doing so, few examine gene inheritance and expression mechanisms. This paper seeks to address this shortcoming by demonstrating how simple changes to these mechanisms can radically alter a population's ability to speciate. Specifically, we find that simulated meiotic inheritance with complete genetic dominance significantly improves a population's divergent capability, even absent other speciation mechanisms.

The ability to learn makes intelligent systems more adaptive. One approach to the development of learning algorithms is to evolve them using evolutionary algorithms. The evolution of learning is interesting as a practical matter because harnessing it may allow us to develop better artificial intelligence; it is also interesting from a theoretical perspective of understanding how the sophisticated learning seen in nature could have arisen. A potential obstacle to the evolution of learning when alternative behavioral strategies (e.g., instincts) can evolve is that learning individuals tend to exhibit ineffective behavior before effective behavior is learned. Nurturing, defined as one individual investing in the development of another individual with which it has an ongoing relationship, is often seen in nature in species that exhibit sophisticated learning behavior. It is hypothesized that nurturing may be able to increase the competitiveness of learning in an evolutionary environment by ameliorating the consequences of incorrect initial behavior. Here we expand upon a foundational work in the evolution of learning to also enable the evolution of instincts and then examine the strategies evolved with and without a nurturing condition in which individuals are not penalized for mistakes made during a learning period. It is found that nurturing promotes the evolution of generalized supervised learning in these environments.

To deal with uncertain environments, autonomous agents must learn without supervision. However, reward-based interactive learning often exhibits limitations handling both generalizations and exceptions, and might benefit from explicit recall of past experiences. For these reasons, this research extends the Context-Aware Learning Model (CALM) with a new learning algorithm. CALM-epLRB combines logistic regression backpropagation in artificial neural networks with hyperbolic reward-based learning and a knowledge base that enables experience-powered learning using a novel weight update rule. CALM-epLRB is compared to two previous CALM versions using four metrics on six data sets and shows promising performance.

To deal with uncertain environments, autonomous agents must learn without supervision. However, reward-based interactive learning often exhibits limitations handling both generalizations and exceptions. For these reasons, we have recently introduced the Context-Aware Learning Model (CALM) and associated learning algorithms. Here we extend CALM by adding a coherent set of neurobiologically inspired mechanisms to produce the neuro-experience-powered version of CALM using logistic-regression backpropagation. CALM-nepLRB is evaluated using four metrics on six data sets and shows promising performance.

An important, long-term objective of intelligent robotics is to develop robots that can learn about and adapt to new environments. We focus on developing a learning model that can build up new knowledge through direct experience with and feedback from an environment. We designed and constructed Context-based Adaptive Robot Behavior-Learning Model (CARB-LM) which is conceptually inspired by Hebbian and anti-Hebbian learning and by neuromodulation in neural networks. CARB-LM has two types of learning processes: (1) context-based learning and (2) reward-based learning. The former uses past accumulated positive experiences as analogies to current conditions, allowing the robot to infer likely rewarding behaviors, and the latter exploits current reward information so the robot can refine its behaviors based on current experience. The reward is acquired by checking the effect of the robot's behavior in the environment. As a first test of this model, we tasked a simulated TurtleBot robot with moving smoothly around a previously unexplored environment. We simulated this environment using ROS and Gazebo and performed experiments to evaluate the model. The robot showed substantial learning and greatly outperformed both a hand-coded controller and a randomly wandering robot.

Speciation is one of the most fundamental and important processes in evolutionary biology, resulting in the panoply of biological diversity found in the natural world. Speciation likewise has profound implications for artificial life, evolutionary computation, and evolutionary robotics, yet a great many aspects of it remain unexplored. Traditionally, speciation was mainly viewed as taking place allopatrically. More recently, sympatric speciation, which does not require geographic isolation, has been studied. Sympatric speciation raises a number of interesting questions with regard to how and why sympatric populations diverge, some of which we address with a 2x2x2 factorial study that considers the factors of sexual selection, resource distribution, and population size. Our hypotheses were evaluated using a synthetic environment inspired by life on the Galápagos Islands. In particular, the wet and dry season dynamics were modeled to produce the intense selection pressure found there. Our results provide direct evidence for the importance of both female mate choice and resource availability on speciation. They also suggest that the greater stability afforded by larger populations can lead to subpopulations between which gene flow is reduced.

Weather radar products from the U.S. National Weather Service (NWS) are used by the government and private sectors. Very high resolution radar data are increasingly being utilized in real time. However, the bandwidth needed to transmit these data (termed level-II super-resolution data) from the radar to the destination site is a limiting issue. General-purpose compression programs are not tuned to the properties of weather radar data. As the NWS continues to upgrade the capabilities of radar network, the amount of data will continue to increase. As a result, compression is of vital interest to keep down maintenance, storage, and transmission costs. A method for lossless compression of these data on a radial-by-radial basis focusing on the delta (difference) between range bins of super-resolution radar data is presented and is called super-resolution delta compression (SRDC). There are several specialized aspects of SRDC that are based on the properties of weather radar data. SRDC was tested on level-II reflectivity product data from several S-band Doppler weather radars in the NWS network and was compared with two general-purpose compression programs and a different weather-radar-specific compression approach. The results show that the newly developed SRDC yield is approximately 17% better than the next best approach and approximately 47% better than only preprocessed radials.

Many complex problems can be solved by a sequence of steps or simple heuristics. In many cases a good solution relies on both good heuristics and proper ordering of their application. Problems such as creating a constrained path through a graph can be solved in this way if we can find a mechanism for ordering the heuristics. We propose using a genetic algorithm (GA) for this purpose to solve the snake-in-the-box problem. However, the additional layer of abstraction created by heuristics weakens the guiding effects of traditional GA operators. We also propose a new set of GA operators that solve this problem by, in the manipulation of the population, applying information from the paths (snakes) produced by the population members. In addition we show the efficacy of this approach by producing a new record length snake of 98 edges in an eight dimensional hypercube. We also compare our new operators with more traditional ones.

A method for identifying values for a genetic algorithm's probability of crossover, mutation rate, and selection pressure that promote the evolution of better results in fewer generations has recently been proposed. This approach, termed the Triple Parameter Hypothesis, uses schema theory to derive these values. However, in initial experimental tests of the hypothesis, the schema distances used were at the extreme ends of the spectrum. In the work presented here, we evaluate the parameters predicted by the hypothesis in a series of maintenance scheduling experiments which use a schema distance in between these extremes. Results show that evolutionary runs which use parameters that satisfy the hypothesis statistically significantly outperform runs that use parameters that do not satisfy the hypothesis, indicating that the hypothesis can be a general purpose process for identifying "good" parameter values

When an optimization problem is encoded using genetic algorithms, one must address issues of population size, crossover and mutation operators and probabilities, stopping criteria, selection operator and pressure, and fitness function to be used in order to solve the problem. This paper tests a relationship between (1) crossover probability, (2) mutation probability, and (3) selection pressure using two problems. This relationship is based on the schema theorem proposed by Holland and reflects the fact that the choice of parameters and operators for genetic algorithms needs to be problem specific.

We propose that abstracting the actions of a behavior coordination mechanism promotes the faster development and higher fitness of an effective controller for complex, composite tasks. Various techniques are well suited for the development of controllers for individual simple tasks. However, as individual tasks are combined into complex, composite tasks, many of these techniques quickly become impractical. By reusing existing behaviors, the focus of development for a controller can be shifted from low-level control to high-level coordination of these existing behaviors. As a result, the development of an effective controller becomes far more practical. Experiments using a single-agent task in a continuous environment demonstrate that grammatical evolution is capable of discovering fuzzy rulesets which effectively coordinate existing behaviors in a controller in fewer generations and with higher fitness than monolithic controllers.

Bloat may occur when evolution allows chromosome growth. Recently it has been shown that elitism can inhibit bloat. Here we study interactions between growth, elitism, and fitness landscapes. Our results show that in some cases elitism neither constrains growth nor increases the rate of fitness accumulation, and when elitism does constrain growth it may stall the search completely.

In evolutionary computation, experimental results are commonly analyzed using an algorithmic performance metric called best-so-far. While best-so-far can be a useful metric, its use is particularly susceptible to three pitfalls: a failure to establish a baseline for comparison, a failure to perform significance testing, and an insufficient sample size. The nature of best-so-far means that it is highly susceptible to these pitfalls. If these pitfalls are not avoided, the use of the best-so-far metric can lead to confusion at best and misleading results at worst. We detail how the use of multiple experimental runs, random search as a baseline, and significance testing can help researchers avoid these common pitfalls. Furthermore, we demonstrate how best-so-far can be an effective algorithmic performance metric if these guidelines are followed.

In problem domains such as robotic control, where the evaluation of an individual significantly dominates the rest of the evolutionary process with respect to time, the viability of an evolutionary approach can be called into question. In an effort to minimize the number of evaluations by maximizing the learning that takes place during an evaluation, a new crossover operator for genetic programming, memetic crossover, was recently introduced. This work analyzes the genealogical impact of this operator at varying levels. Although diversity, both in terms of individuals and nodes, is reduced in memetic crossover, we show that memetic crossover is capable of working with standard crossover to consistently improve the fitness of individuals.

The combination of fuzzy control and behavior hierarchies allows for the construction of more complex behavior-based robot control agents than does either technique alone. However, current implementations are limited in their complexity since high-level behaviors still use low-level sensor information. We propose a technique for abstracting this low-level sensor information into priorities which are used to completely abstract out the context in which a high-level, fuzzy behavior operates. This modification enables a single high-level behavior to coordinate the lower-level behaviors within a single robot, among the robots in a team, or even among teams of teams. This is demonstrated in a scenario in which multiple pursuers attempt to capture a prey.

With applications in coding theory and hypercube-based computing and networking, the in the box problem is of great practical importance. Moreover, it is both mathematically elegant and highly difficult. The problem is simply to find the longest in a hypercube. However, as the hypercube grows in dimensionality, the size of the search space increases exponentially. Moreover, as the maximum snake length is only known for the smallest dimensions (where the snakes themselves have already been identified), there is no known stopping criterion for the search in higher dimensions. In this paper we make a mathematical conjecture about the possible maximum length of a snake in a hypercube of dimension d. We use a genetic algorithm for finding snakes in a 8 hypercube to show some results.

Hunting for snakes of maximum length in hypercubes has been addressed with non-heuristic methods for hypercubes of dimension less than eight. Above that dimension the problem is intractable because the search grows exponentially with the dimension, which make it an NP-hard problem. Heuristic methods, like genetic algorithms, have been used to solve this kind of problem. We propose different fitness functions to find snakes in hypercubes of dimension greater than three and pose some open questions regarding the number of maximum length snakes in a hypercube of dimension d

Convergence to correct solutions in Genetic Algorithms depends largely on the fitness function. A fitness function that captures all goals and constraints can be difficult to find. This paper gives a mathematical justification for a fitness function that has previously been demonstrated experimentally to be effective.

For some problem domains, the evaluation of individuals is significantly more expensive than the other steps in the evolutionary process. Minimizing these evaluations is vital if we want to make genetic programming a viable strategy. In order to minimize the required evaluations, we need to maximize the amount learned from each evaluation. To accomplish this, we introduce a new crossover operator for genetic programming, memetic crossover that allows individuals to imitate the observed success of others. An individual that has done poorly in some parts of the problem may then imitate an individual that did well on those same parts. This results in an intelligent search of the feature-space, and therefore fewer evaluations.

We have designed and built a set of miniature robots called Scouts and have developed a distributed software system to control them. This paper addresses the fundamental choices we made in the design of the control software, describes experimental results in a surveillance task, and analyzes the factors that affect robot performance. Space and power limitations on the Scouts severely restrict the computational power of their on-board computers, requiring a proxy-processing scheme in which the robots depend on remote computers for their computing needs. While this allows the robots to be autonomous, the fact that robots' behaviors are executed remotely introduces an additional complication-sensor data and motion commands have to be exchanged using wireless communications channels. Communications channels cannot always be shared, thus requiring the robots to obtain exclusive access to them. We present experimental results on a surveillance task in which multiple robots patrol an area and watch for motion. We discuss how the limited communications bandwidth affects robot performance in accomplishing the task, and analyze how performance depends on the number of robots that share the bandwidth.

We describe a distributed software system for controlling a group of miniature robots using a very low capacity communication system. Space and power limitations on the miniature robots drastically restrict the capacity of the communication system and require sharing bandwidth and other resources among the robots. We have developed a process management/scheduling system that dynamically assigns resources to each robot in an attempt to maximize the utilization of the available resources while still maintaining a priori behavior priorities. We describe a surveillance task in which the robots patrol an area and watch for motion, and present experimental results.

Presents a miniature robotic system ("scout") useful for reconnaissance and surveillance missions. A large number of scout robots are deployed and controlled by humans and/or larger "ranger" robots. The specially designed and constructed scouts are extremely small (roughly 116cc volume) yet are readily deployable (by tossing or launching), have multiple mobility modes, have multiple sensing capabilities, can transmit and receive data and instructions, and have a limited capability for autonomous action. The rangers are significantly larger vehicles, based on a commercial-off-the-shelf platform, augmented with scout launchers, radios, and additional sensors. Together, the scouts and rangers form a hierarchical team capable of carrying out complex missions in a wide variety of environments.

Distributed robotics uses multiple robots through coordination of task to solve problems. Developing a distributed robotic system introduces a number of challenges, one of which is the intra-robot communication and networking of data and commands. Due to the characteristics of the targeted robotic missions and environments, the wireless networking requirements impose stringent constraints in terms of available memory, processing, and power consumption. Designing under these constraints and still providing support for such functions as synchronous commands and proxy processing is the true challenge. This paper discusses research in providing a lightweight wireless routing/networking solution that can dynamically tune the intra-robotic networking to adapt to the robotic team's mission needs and the environment situation encountered.

This paper presents an application of a connectionist control-learning system designed for use on an autonomous mini-robot. This system was formerly shown to form useful two-dimensional mappings rapidly when applied to backing a car with a single trailer. In the current paper the learning system is extended to three dimensions and applied to a similar but significantly more difficult problem. The system is shown to be capable of rapid unsupervised learning of output responses in temporal domains through the use of eligibility traces and inter-neural cooperation within topologically defined neighborhoods.

This paper presents the application of a connectionist control-learning system to an autonomous mini-robot. The system's design is severely constrained by the computing power and memory available on board the mini-robot and the on-board training time is greatly limited by the short life of the battery. The system is capable of rapid unsupervised learning of output responses in temporal domains through the use of eligibility traces and data sharing within topologically defined neighborhoods.