Large Language Models (LLMs) have been explored for automating or enhancing penetration testing tasks, but their effectiveness and reliability across diverse attack phases remain open questions.This study presents a comprehensive evaluation of multiple LLM-based agents, ranging from singular to modular designs, across realistic penetration testing scenarios, analyzing their empirical performance and recurring failure patterns.We further investigate the impact of core functional capabilities on agent success, operationalized through five targeted augmentations: Global Context Memory (GCM), Inter-Agent Messaging (IAM), Context-Conditioned Invocation (CCI), Adaptive Planning (AP), and Real-Time Monitoring (RTM).These interventions respectively support the capabilities of Context Coherence & Retention, Inter-Component Coordination & State Management, Tool Usage Accuracy & Selective Execution, Multi-Step Strategic Planning & Error Detection & Recovery, and Real-Time Dynamic Responsiveness.Our findings reveal that while some architectures natively exhibit select properties, targeted augmentations significantly enhance modular agent performance-particularly in complex, multi-step, and real-time penetration testing scenarios.

OBJECTIVE: Identifying fraud in healthcare programs is crucial, as an estimated 3%-10% of the total healthcare expenditures are lost to fraudulent activities. This study presents a systematic literature review of machine learning techniques applied to fraud detection in health insurance claims. We aim to analyze the data and methodologies documented in the literature over the past two decades, providing insights into research challenges and opportunities. METHODS: We identified research studies on health insurance fraud detection using machine learning approaches from databases such as Google Scholar, Springer-Link journals, Elsevier, PubMed, Excerpta Medica Database (EMBASE), Scopus, the Association for Computing Machinery (ACM) Digital Library, and the Institute of Electrical and Electronics Engineers (IEEE) Xplore Digital Library. We included only articles that presented experimental results of machine learning-based approaches applied to healthcare claims. From the reviewed articles, 137 were selected for the final qualitative and quantitative analyses. RESULTS: In recent years, there has been a surge in publications centered on the use of machine learning to detect health insurance fraud. Among these studies, those focused on the detection of fraud committed by healthcare providers was the most prevalent, followed by fraud committed by patients. A wide variety of machine learning algorithms are highlighted in these studies, ranging from unsupervised (41 studies) and supervised methods (94 studies), to hybrid approaches (12 studies). While traditional machine learning approaches remain dominant in this research area, the adoption of advanced deep learning techniques is on the rise. Considering the type of healthcare claims data used, 30 studies utilized private data sources, while the rest used publicly available datasets. Data from 16 countries were utilized, with a majority coming from the United States (96 studies), followed by China (11 studies) and Australia (5 studies). DISCUSSION AND CONCLUSION: Detecting fraud in healthcare claims using machine learning presents several challenges. These include inconsistent data, absence of data standardization and integration, privacy concerns, and a limited number of labeled fraudulent cases to train models on. Future work should focus on enhancing transparency in data preparation, promoting the sharing of fraud investigation outcomes by authorities, and developing benchmark datasets to enhance accessibility and comparability. Furthermore, innovative techniques in data sampling, feature encoding methods for training machine learning models, and exploring the latest advancements in deep learning can significantly advance research in health insurance fraud detection.

In the burgeoning field of human-AI interaction, trust emerges as a cornerstone because many think that it is critical to the effectiveness of collaboration and the acceptance of AI systems. Traditional methods of assessing trust have predominantly relied on self-reported measures, requiring participants to articulate their perceptions and attitudes through questionnaires. However, these explicit methods may not fully capture the nuanced dynamics of trust, especially in real-time and complex interaction environments. This paper introduces an innovative approach to evaluating trust in human-AI teams, pivoting from the conventional reliance on verbal or written feedback to analyzing gameplay behaviors as implicit indicators of trust levels. Utilizing the Overcooked-AI environment, our study explores how participants’ interactions with AI agents of varying performance levels can reveal underlying trust mechanisms without a single query posed to the human players. This approach not only bypasses the efficiency challenges posed by repetitive and lengthy trust assessment methods, but also provides insights comparable to them. We highlight the potential of non-verbal cues and action patterns as reliable trust indicators by comparing the predictive accuracies of questionnaire-based models with those derived from gameplay behavior analysis. Furthermore, our findings suggest that these implicit measures can be integrated into adaptive systems and algorithms for real-time trust calibration in human-agent teaming settings. This shift towards an action-oriented trust assessment challenges existing paradigms and opens new avenues for understanding and enhancing human-AI collaboration.

While machine learning models have demonstrated the ability to detect cyber attacks, their deployment to operational scenarios is often limited due to the possibility of the model failing to detect attacks not included in the training set. Transfer learning has been shown to be a possible response to this common problem in the cyber domain. However, there are many operational questions that must be addressed before transfer learning can be implemented. First, and foremost, is the question of how to detect a change in an operational environment (e.g., the presence of a new attack variant). Transfer distance is a measure of dissimilarity between two learning problems. This study evaluates common distance metrics that could be used to estimate transfer distance between cyber attack learning problems, proposes a new method for calculating a multivariate transfer distance measure, and formulates a method for monitoring streaming data for changes in the environment using a one class Naïve Bayes model. These techniques are tested and evaluated on a publicly available cyber attack detection data set that contains multiple attack types.

Reinforcement learning (RL) operating on attack graphs leveraging cyber terrain principles are used to develop reward and state associated with determination of surveillance detection routes (SDR). This work extends previous efforts on developing RL methods for path analysis within enterprise networks. This work focuses on building SDR where the routes focus on exploring the network services while trying to evade risk. RL is utilized to support the development of these routes by building a reward mechanism that would help in realization of these paths. The RL algorithm is modified to have a novel warm-up phase which decides in the initial exploration which areas of the network are safe to explore based on the rewards and penalty scale factor.

Active learning is a practical field of machine learning that automates the process of selecting which data to label. Current methods are effective in reducing the burden of data labeling but are heavily model-reliant. This has led to the inability of sampled data to be transferred to new models as well as issues with sampling bias. Both issues are of crucial concern in machine learning deployment. We propose active learning methods utilizing combinatorial coverage to overcome these issues. The proposed methods are data-centric, as opposed to model-centric, and through our experiments we show that the inclusion of coverage in active learning leads to sampling data that tends to be the best in transferring to better performing models and has a competitive sampling bias compared to benchmark methods.

Abstract Cyber-physical systems are complex systems that require the integration of diverse software, firmware, and hardware to be practical and useful. This increased complexity is impacting the management of models necessary for designing cyber-physical systems that are able to take into account a number of “-ilities”, such that they are safe and secure and ultimately resilient to disruption of service. We propose an ontological metamodel for system design that augments an already existing industry metamodel to capture the relationships between various model elements (requirements, interfaces, physical, and functional) and safety, security, and resilient considerations. Employing this metamodel leads to more cohesive and structured modeling efforts with an overall increase in scalability, usability, and unification of already existing models. In turn, this leads to a mission-oriented perspective in designing security defenses and resilience mechanisms to combat undesirable behaviors. We illustrate this metamodel in an open-source GraphQL implementation, which can interface with a number of modeling languages. We support our proposed metamodel with a detailed demonstration using an oil and gas pipeline model.

The exploitation of extra state information has been an active research area in multi-agent reinforcement learning (MARL). QMIX represents the joint action-value using a non-negative function approximator and achieves the best performance on the StarCraft II micromanagement testbed, a common MARL benchmark. However, our experiments demonstrate that, in some cases, QMIX performs sub-optimally with the A2C framework, a training paradigm that promotes algorithm training efficiency. To obtain a reasonable trade-off between training efficiency and algorithm performance, we extend value-decomposition to actor-critic methods that are compatible with A2C and propose a novel actor-critic framework, value-decomposition actor-critic (VDAC). We evaluate VDAC on the StarCraft II micromanagement task and demonstrate that the proposed framework improves median performance over other actor-critic methods. Furthermore, we use a set of ablation experiments to identify the key factors that contribute to the performance of VDAC.

Metric learning is a well-developed field in machine learning and has seen recent application in the area of prognostics and health management (PHM). Metric learning allows for fault diagnosis or condition monitoring models to be developed with the assumption that a machine- or load-specific similarity metric can be learned after model deployment. Existing literature has used metric learning to fine-tune deep learning models to address machine-to-machine differences and differences in working conditions. Here, we study metric learning in isolation, not as an intermediate step in deep learning, by conducting a comparative study of Principal Component Analysis (PCA), Neighborhood Component Analysis (NCA), Local Fisher Discriminant Analysis (LFDA), and Large Margin Nearest Neighbor (LMNN). We consider performance metrics for prediction performance, cluster performance, feature sensitivity, sample efficiency, and latent space efficiency. We find that linear partitions on the latent spaces learned via metric learning are able to achieve accuracies greater than 90% on Case Western Reserve University’s bearing fault data set using only the drive-end vibration signal. We find PCA to be dominated by metric learning algorithms for all working loads considered. And, in sum, we demonstrate classical metric learning algorithms to be a promising approach for learning machine-and load-specific similarity metrics for PHM with minor data processing and small samples.

Supervised learning algorithms require a set of labeled training data. In many engineering applications, acquiring and accurately labeling the training data can be time consuming, burdensome, and costly. Active learning is an area of machine learning that selects observations in an unlabeled set to be passed to an oracle to retrieve the ground truth label, thereby improving the efficiency of the labeling and training process. However, most active learning algorithms only consider model improvement and, therefore, ignore cost considerations. The active learning algorithms that do consider cost require the practitioner to specify the trade-off between model improvement and cost. We propose an active learning with cost method that does not require this trade-off to be specified by randomly sampling observations from the Pareto optimal frontier. Further, we propose an extension to this method that accounts for uncertainty in the cost estimate of labeling an observation. These methods are evaluated on publicly available data sets, and the numerical experiments demonstrate that the proposed methods can produce models that achieve similar performance to standard active learning algorithms while reducing the labeling cost.

In adversarial environments, one side could gain an advantage by identifying the opponent’s strategy. For example, in combat games, if an opponent’s strategy is identified as overly aggressive, one could lay a trap that exploits the opponent’s aggressive nature. However, an opponent’s strategy is not always apparent and may need to be estimated from observations of their actions. This paper proposes to use inverse reinforcement learning (IRL) to identify strategies in adversarial environments. Specifically, the contributions of this work are 1) the demonstration of this concept on gaming combat data generated from three pre-defined strategies and 2) the framework for using IRL to achieve strategy identification. The numerical experiments demonstrate that the recovered rewards can be identified using a variety of techniques including visual analysis, cluster analysis, and supervised classification.

In this article, we propose an end-to-end adaptive framework for optimal trade execution based on Proximal Policy Optimization (PPO). We use two methods to account for the time dependencies in the market data based on two different neural network architecture: 1) Long short-term memory (LSTM) networks, 2) Fully-connected networks (FCN) by stacking the most recent limit orderbook (LOB) information as model inputs. The proposed framework can make trade execution decisions based on level-2 limit order book (LOB) information such as bid/ask prices and volumes directly without manually designed attributes as in previous research. Furthermore, we use a sparse reward function, which gives the agent reward signals at the end of each episode as an indicator of its relative performances against the baseline model, rather than implementation shortfall (IS) or a shaped reward function. The experimental results have demonstrated advantages over IS and the shaped reward function in terms of performance and simplicity. The proposed framework has outperformed the industry commonly used baseline models such as TWAP, VWAP, and AC as well as several Deep Reinforcement Learning (DRL) models on most of the 14 US equities in our experiments.

Static analysis tools are programs that run on source code prior to their compilation to binary executables and attempt to find flaws or defects in the code during the early stages of development. If left unresolved, these flaws could pose security risks. While numerous static analysis tools exist, there is no single tool that is optimal. Therefore, many static analysis tools are often used to analyze code. Further, some of the alerts generated by the static analysis tools are low-priority or false alarms. Machine learning algorithms have been developed to distinguish between true alerts and false alarms, however significant man hours need to be dedicated to labeling data sets for training. This study investigates the use of active learning to reduce the number of labeled alerts needed to adequately train a classifier. The numerical experiments demonstrate that a query by committee active learning algorithm can be utilized to significantly reduce the number of labeled alerts needed to achieve similar performance as a classifier trained on a data set of nearly 60,000 labeled alerts.

This paper addresses the problem of multi-agent inverse reinforcement learning (MIRL) in a two-player general-sum stochastic game framework. Five variants of MIRL are considered: uCS-MIRL, advE-MIRL, cooE-MIRL, uCE-MIRL, and uNE-MIRL, each distinguished by its solution concept. Problem uCS-MIRL is a cooperative game in which the agents employ cooperative strategies that aim to maximize the total game value. In problem uCE-MIRL, agents are assumed to follow strategies that constitute a correlated equilibrium while maximizing total game value. Problem uNE-MIRL is similar to uCE-MIRL in total game value maximization, but it is assumed that the agents are playing a Nash equilibrium. Problems advE-MIRL and cooE-MIRL assume agents are playing an adversarial equilibrium and a coordination equilibrium, respectively. We propose novel approaches to address these five problems under the assumption that the game observer either knows or is able to accurately estimate the policies and solution concepts for players. For uCS-MIRL, we first develop a characteristic set of solutions ensuring that the observed bi-policy is a uCS and then apply a Bayesian inverse learning method. For uCE-MIRL, we develop a linear programming problem subject to constraints that define necessary and sufficient conditions for the observed policies to be correlated equilibria. The objective is to choose a solution that not only minimizes the total game value difference between the observed bi-policy and a local uCS, but also maximizes the scale of the solution. We apply a similar treatment to the problem of uNE-MIRL. The remaining two problems can be solved efficiently by taking advantage of solution uniqueness and setting up a convex optimization problem. Results are validated on various benchmark grid-world games.

The focus of this paper is a Bayesian framework for solving a class of problems termed multiagent inverse reinforcement learning (MIRL). Compared to the well-known inverse reinforcement learning (IRL) problem, MIRL is formalized in the context of stochastic games, which generalize Markov decision processes to game theoretic scenarios. We establish a theoretical foundation for competitive two-agent zero-sum MIRL problems and propose a Bayesian solution approach in which the generative model is based on an assumption that the two agents follow a minimax bipolicy. Numerical results are presented comparing the Bayesian MIRL method with two existing methods in the context of an abstract soccer game. Investigation centers on relationships between the extent of prior information and the quality of learned rewards. Results suggest that covariance structure is more important than mean value in reward priors.

With the emergence of the Microsoft Kinect sensor, many developer communities and research groups have found countless uses and have already published a wide variety of papers that utilize the raw depth images for their specific goals. New methods and applications that use the device generally require an appropriately large ensemble of data sets with accompanying ground truth for testing purposes, as well as accurate models that account for the various systematic and stochastic contributors to Kinect errors. Current error models, however, overlook the intermediate infrared (IR) images that directly contribute to noisy depth estimates. We, therefore, propose a high fidelity Kinect IR and depth image predictor and simulator that models the physics of the transmitter/receiver system, unique IR dot pattern, disparity/depth processing technology, and random intensity speckle and IR noise in the detectors. The model accounts for important characteristics of Kinect's stereo triangulation system, including depth shadowing, IR dot splitting, spreading, and occlusions, correlation-based disparity estimation between windows of measured and reference IR images, and subpixel refinement. Results show that the simulator accurately produces axial depth error from imaged flat surfaces with various tilt angles, as well as the bias and standard lateral error of an object's horizontal and vertical edge.

Electronic markets and automated trading have resulted in a drastic increase in the quantity and complexity of regulatory data. Regulatory analysis now includes detailed analysis of all messaging and communications related to electronic limit order books. New order types, intra-market behavior and other exchange functionality further complicate analysis. Data visualizations have proven to be a fundamental tool for building intuition and enabling exploratory data analysis in many fields. In this paper, we propose the incorporation of visualizations in the workflow of multiple financial regulatory roles, including market surveillance, enforcement, and academic research.

Limit order book simulations based on “zero-intelligence” or “entropy-maximizing” agents address two difficult issues in financial economics. First, the models address the significance of trading mechanisms by explicitly accounting for the logic of those mechanisms. Second, they avoid the difficulty of modeling human decision-making by generating orders stochastically. This paper reports on a computational experiment in which a strategic agent trading on endogenous market signals is embedded in an otherwise stochastic order book simulation. Under certain parameterizations of the model the agent is profitable despite the fact that the agent only employs market orders.

Noise characteristics for the Microsoft Kinect sensor are presented. Horizontal (x) and vertical (y) stochastic noise are measured using a novel 3D checker board. Results show that the noise is affected mostly by the depth at which the object is sensed and by the radial distance from the center of the field of view. Measurement-based models for the noise in horizontal and vertical axes are presented. The proposed model is compared against existing models in literature and shows better results by considering the horizontal and vertical location of the depth measurement.

Since the emergence of the commodity-priced, small-sized Microsoft Kinect™ sensor, 3D object pose estimation has become prevalent in many applications across a wide variety of disciplines. However, because most current methods require a hard assignment between a measurement point cloud and a given CAD model point cloud for alignment, accurate pose estimation is limited to a small range of sensor noise and resolution, and CAD model precision. This paper therefore presents a MLE algorithm that achieves a statistically optimal global maximum likelihood on the surface of the continuous 6D pose domain by soft assigning all measurement points to all model points. The accuracy in estimation of orientation and position of the MLE algorithm is then compared to a variant of the ICP method that accounts for an anisotropie Gaussian measurement noise distribution. It is finally shown through a series of simulated measurement point clouds and depth images that MLE outperforms the ICP variant, and achieves a monotonie increase in performance with an increase in either points on target or CAD model precision.

Identifying and understanding the impact of algorithmic trading on financial markets has become a critical issue for market operators and regulators. Advanced data feed and audit trail information from market operators now make the full observation of market participants' actions possible. A key question is the extent to which it is possible to understand and characterize the behavior of individual participants from observations of trading actions. In this paper, we consider the basic problems of categorizing and recognizing traders (or, equivalently, trading algorithms) on the basis observed limit orders. Our approach, which is based on inverse reinforcement learning (IRL), is to model trading decisions as a Markov decision process and then use observations of an optimal decision policy to find the reward function. The approach strikes a balance between two desirable features in that it captures key empirical properties of order book dynamics and yet remains computationally tractable. Making use of a real-world data set from the E-Mini futures contract, we compare two principal IRL variants, linear IRL and Gaussian process IRL. Results suggest that IRL-based feature spaces support accurate classification and meaningful clustering.

We propose a zero-intelligence agent-based model of the E-Mini S&P 500 futures market, which allows for a close examination of the market microstructure. Several classes of agents are characterized by their order speed and order placement within the limit order book. These agents' orders populate the simulated market in a way consistent with real world participation rates. By modeling separate trading classes the simulation is able to capture interactions between classes, which are essential to recreating market phenomenon. The simulated market is validated against empirically observed characteristics of price returns and volatility. We therefore conclude that our agent based simulation model can accurately capture the key characteristics of the nearest months E-Mini S&P 500 futures market. Additionally, to illustrate the applicability of the simulation, experiments were run, which confirm the leading hypothesis for the cause of the May 6th2010 Flash Crash.

Electronic markets have emerged as popular venues for the trading of a wide variety of financial assets, and computer based algorithmic trading has also asserted itself as a dominant force in financial markets across the world. Identifying and understanding the impact of algorithmic trading on financial markets has become a critical issue for market operators and regulators. We propose to characterize traders' behavior in terms of the reward functions most likely to have given rise to the observed trading actions. Our approach is to model trading decisions as a Markov Decision Process (MDP), and use observations of an optimal decision policy to find the reward function. This is known as Inverse Reinforcement Learning (IRL), and a variety of approaches for this problem are known. Our IRL-based approach to characterizing trader behavior strikes a balance between two desirable features in that it captures key empirical properties of order book dynamics and yet remains computationally tractable. Using an IRL algorithm based on linear programming, we are able to achieve more than 90% classification accuracy in distinguishing High Frequency Trading from other trading strategies in experiments on a simulated E-Mini S&P 500 futures market. The results of these empirical tests suggest that High Frequency Trading strategies can be accurately identified and profiled based on observations of individual trading actions.

We consider decentralized Bayesian search problems that involve a team of multiple autonomous agents searching for targets on a network of search points operating under the following constraints: 1) interagent communication is limited; 2) the agents do not have the opportunity to agree in advance on how to resolve equivalent but incompatible strategies; and 3) each agent lacks the ability to control or predict with certainty the actions of the other agents. We formulate the multiagent search-path-planning problem as a decentralized optimal control problem and introduce approximate dynamic heuristics that can be implemented in a decentralized fashion. After establishing some analytical properties of the heuristics, we present computational results for a search problem involving two agents on a 5 x 5 grid.

To quantify text-based unstructured information, we propose a method called the direct scoring algorithm (DSA). DSA uses keywords in the document, subjectively-determined numerical weights, and subjectively-designed grammar rules to score individual sentences. We use our methods to score the Beige books produced by the U.S. Federal Reserve, which contain subjective text-based commentary on state of the economy. To assess whether our scores have value in a predictive sense, we use them to construct a linear regression model of future growth in U.S. gross domestic product (GDP). We then compare the performance characteristics of this model with those a similar model based on scores of the same documents produced though subjective reading by professional economists. The comparison demonstrates that the DSA model using the Beige book significantly contributes to the prediction of GDP growth, explaining as much as 69% of the variance compared to the scores created by economic experts. We also add the extracted section scores to a GDP time series prediction model, which uses only structured data as input. The results of this experiment suggest the unstructured information in the Beige books has predictive value that goes beyond that of the structure information used in the time series model, and that our approach has some potential as a means of extracting this information in a semi-automated fashion.

Feature selection is a significant problem in building any predictive model. Linear programming models which minimize the sum of deviations reward addition of variables if the added variables can reduce the sum of deviations. Deviation occurs when a point falls on the wrong side of the discriminant surface. If the groups are not linearly separable, and if the number of features is large, it is possible to create a model where some of the features used in the model account for a very small reduction in the deviations. We propose a feature selection scheme for LP models in which we measure the effect of each variable in increasing the interclass separation.

Multi-modal problems are amongst the most difficult-to-handle classification problems, especially for traditional statistical techniques. Multi-modal problems arise when each class region can occupy disjoint areas in feature space. Backpropagation neural networks and decision tree classifiers (DTCs) can typically handle multi-modal problems. We introduce a decision tree based on clustering and linear programming and compare its performance to CART on a number of data sets from the literature, including several sets that exhibit clear multi-modal structure.

Most existing approaches to learning the structure of Bayesian networks assume that all variables are discrete or that all variables are continuously normally distributed. We propose a meta-Gaussian approach that is appropriate for direct learning from general, continuous variables. We first transform the original variables into standard normal variables. Under the assumption that the transformed variables are multivariate normally distributed, we then make use of existing algorithms to learn the network structure in the transformed space, and then project the results back into the original space. Preliminary experimental results show that this approach can recover the network structure, provided that the variables of the network satisfy a fundamental monotonicity property.

In an earlier paper, the authors presented a new parameterization algorithm for the the linear complementarity problem. The trajectory associated with this parameterization is distinguished by a naturally defined starting point and by a piecewise characterization as a fractional polynomial function of a single parameter. In order to follow this trajectory, however one needs to isolate roots of polynomials-a computationally expensive operation. In this paper, we present an algorithm in which we parameterize one dimension at a time. This results in polynomials of degree one, which allows trivial root isolation. The algorithm stays unaffected in other key respects. For example, the average number of pieces in the trajectory is still O(n/sup 2/), where n is the dimension of the problem space. This implies that our algorithm is competitive, in an average sense, to Lemke's method.

Among the various parametric and nonparametric techniques available for classification, k-nearest neighbor is a well known nonparametric method. This paper describes some experiments with modification of a k-nearest neighbor method to use the neighbor information as attributes to a second k-nearest neighbor classifier.

We investigate the performance of a linear programming-based decision tree in creating gray scale images for road extraction from AVIRIS images. We apply our method for classification of pixels from a digital image of an area near Williamsburg, Virginia, using the distance from discriminant lines as a measure to create a gray scale image. Our method effectively captures information from a large number of bands of the original image and can be a useful input to other techniques which can use only a single band.

We consider the problem that many portfolio managers face of selecting, on a regular basis, stocks for investment and recommendation to clients. In typical solution strategies, binary rules are developed to classify stocks as strong or weak performers based on technical indicators. Strategies based on binary classification rules have been shown to be very effective at maximizing the total profitability of the stocks that are selected. Having a fixed number of target stocks is important for portfolio maintenance and for client choice, however, and so the selection problem also engenders the additional constraint of limiting the total number of stocks selected. Binary classification rule strategies do not address this constraint. In this paper we investigate the use of scoring functions, which have the advantage of allowing one to rank order the population based on profitability, as an alternative to binary classification rules. A key feature of this work is that we develop the scoring functions by incorporating binary classification rules. In particular, we induce the score model by assigning optimal weights to sets of implicit positive binary classification rules. We use a genetic algorithm with supervised batch learning to evolve classification rules. Fitness of a rule set is evaluated based on the success of the scoring function that it induces. We report on the relative empirical performance of this method on several large historical data sets.

We examine the use of numeric score functions that allow one to rank order a universe of stocks based on profitability. We use a genetic algorithm to evolve sets of 'implicit-positive' binary classification rules. Using each rule set, we induce a scoring model by weighting the individual terms in a representation of the rule in terms of binary variables. We report on the empirical performance of the proposed family of scoring algorithms on several large historical stock data sets. We also compare our approach with a polynomial network technique.

A basic hierarchical network design problem is that of selecting access area and backbone designs that minimize the sum total cost of the network. Because of its computational difficulty, network designers typically segment the hierarchical design problem, first solving the access area problem to obtain a set of backbone nodes and then solving the backbone design problem on the subgraph induced by these nodes. Each individual problem is far easier to solve than the complete network design problem, but in general the procedure gives a poor overall solution. In this paper, we describe a technique for integrating the access area and backbone design problems into a single mathematical program. The fundamental idea of this approach is to incorporate backbone network cost information into the access area problem without increasing the computational difficulty of the resulting problem significantly beyond that of the access area problem.

We derive an algorithm based on the ellipsoid method that solves linear programs whose coefficients are real algebraic numbers. By defining the encoding size of an algebraic number to be the bit size of the coefficients of its minimal polynomial, we prove the algorithm runs in time polynomial in the dimension of the problem, the encoding size of the input coefficients, and the degree of any algebraic extension which contains the input coefficients. This bound holds even if all input and arithmetic is performed symbolically, using rational numbers only.

It is shown that a modified variant of the interior point method can solve linear programs (LPs) whose coefficients are real numbers from a subring of the algebraic integers. By defining the encoding size of such numbers to be the bit size of the integers that represent them in the subring, it is proved that the modified algorithm runs in time polynomial in the encoding size of the input coefficients, the dimension of the problem, and the order of the subring. The Tardos scheme is then extended to this case, yielding a running time that is independent of the objective and right-hand side data. As a consequence of these results, it is shown that LPs with real circulant coefficient matrices can be solved in strongly polynomial time. It is also shown how the algorithm can be applied to LPs whose coefficients belong to the extension of the integers by a fixed set of square roots.>