The detection of, explanation of, and accommodation to anomalies and novelties are active research areas in multiple communities, including data mining, machine learning, and computer vision. They are applied in various guises including anomaly detection, out-of-distribution example detection, adversarial example recognition and detection, curiosity-driven reinforcement learning, and open-set recognition and adaptation, all of which are of great interest to the SIGKDD community. The techniques developed have been applied in a wide range of domains including fraud detection and anti-money laundering in fintech, early disease detection, intrusion detection in large-scale computer networks and data centers, defending AI systems from adversarial attacks, and in improving the practicality of agents through overcoming the closed-world assumption.

Open category detection is the problem of detecting “alien" test instances that belong to categories or classes that were not present in the training data. In many applications, reliably detecting such aliens is central to ensuring the safety and accuracy of test set predictions. Unfortunately, there are no algorithms that provide theoretical guarantees on their ability to detect aliens under general assumptions. Further, while there are algorithms for open category detection, there are few empirical results that directly report alien detection rates. Thus, there are significant theoretical and empirical gaps in our understanding of open category detection. In this paper, we take a step toward addressing this gap by studying a simple, but practically-relevant variant of open category detection. In our setting, we are provided with a “clean" training set that contains only the target categories of interest and an unlabeled “contaminated” training set that contains a fraction $\alpha$ of alien examples. Under the assumption that we know an upper bound on $\alpha$, we develop an algorithm that gives PAC-style guarantees on the alien detection rate, while aiming to minimize false alarms. Given an overall budget on the amount of training data, we also derive the optimal allocation of samples between the mixture and the clean data sets. Experiments on synthetic and standard benchmark datasets evaluate the regimes in which the algorithm can be effective and provide a baseline for further advancements. In addition, for the situation when an upper bound for $\alpha$ is not available, we employ nine different anomaly proportion estimators, and run experiments on both synthetic and standard benchmark data sets to compare their performance.

In biodiversity conservation, adaptive management (AM) is the principal tool for decision making under uncertainty. AM problems are planning problems that can be modelled using Mixed Observability MDPs (MOMDPs). MOMDPs tackle decision problems where state variables are completely or partially observable. Unfortunately, MOMDP solutions (policy graphs) are too complex to be interpreted by human decision-makers. Here, we provide algorithms to solve K-N-MOMDPs, where K represents the maximum number of fully observable states and N represents the maximum number of alpha-vectors. Our algorithms calculate compact and more interpretable policy graphs from existing MOMDP models and solutions. We apply these algorithms to two computational sustainability applications: optimal release of bio-control agents to prevent dengue epidemics and conservation of the threatened bird species Gouldian finch. The methods dramatically reduce the number of states and alpha-vectors in MOMDP problems without significantly reducing their quality. The resulting policies have small policy graphs (4-6 nodes) that can be easily interpreted by human decision-makers.

Existing calibration algorithms address the problem of covariate shift via unsupervised domain adaptation. However, these methods suffer from the following limitations: 1) they require unlabeled data from the target domain, which may not be available at the stage of calibration in real-world applications and 2) their performance depends heavily on the disparity between the distributions of the source and target domains. To address these two limitations, we present novel calibration solutions via domain generalization. Our core idea is to leverage multiple calibration domains to reduce the effective distribution disparity between the target and calibration domains for improved calibration transfer without needing any data from the target domain. We provide theoretical justification and empirical experimental results to demonstrate the effectiveness of our proposed algorithms. Compared against state-of-the-art calibration methods designed for domain adaptation, we observe a decrease of 8.86 percentage points in expected calibration error or, equivalently, an increase of 35 percentage points in improvement ratio for multi-class classification on the Office-Home dataset.

The detection of, explanation of, and accommodation to anomalies and novelties are active research areas in multiple communities, including data mining, machine learning, and computer vision. They are applied in various guises including anomaly detection, out-of-distribution example detection, adversarial example recognition and detection, curiosity-driven reinforcement learning, and open-set recognition and adaptation, all of which are of great interest to the SIGKDD community. The techniques developed have been applied in a wide range of domains including fraud detection and anti-money laundering in fintech, early disease detection, intrusion detection in large-scale computer networks and data centers, defending AI systems from adversarial attacks, and in improving the practicality of agents through overcoming the closed-world assumption.

Deep learning approaches to anomaly detection (AD) have recently improved the state of the art in detection performance on complex data sets, such as large collections of images or text. These results have sparked a renewed interest in the AD problem and led to the introduction of a great variety of new methods. With the emergence of numerous such methods, including approaches based on generative models, one-class classification, and reconstruction, there is a growing need to bring methods of this field into a systematic and unified perspective. In this review, we aim to identify the common underlying principles and the assumptions that are often made implicitly by various methods. In particular, we draw connections between classic “shallow” and novel deep approaches and show how this relation might cross-fertilize or extend both directions. We further provide an empirical assessment of major existing methods that are enriched by the use of recent explainability techniques and present specific worked-through examples together with practical advice. Finally, we outline critical open challenges and identify specific paths for future research in AD.

Rainfall is a very important weather variable, especially for agriculture. Unfortunately, rain gauges fail frequently. This paper describes a conditional mixture model for predicting the presence and amount of rain at a weather station based on measurements at nearby stations. The model is evaluated on simulated faults (blocked rain gauges) inserted into observations from the Oklahoma Mesonet. Using the negative log-likelihood as an anomaly score, we evaluate the area under the ROC and precision-recall curves for detecting these faults. The results show very good performance.

Unsupervised anomaly detection algorithms search for outliers and then predict that these outliers are the anomalies. When deployed, however, these algorithms are often criticized for high false-positive and high false-negative rates. One main cause of poor performance is that not all outliers are anomalies and not all anomalies are outliers. In this article, we describe the Active Anomaly Discovery (AAD) algorithm, which incorporates feedback from an expert user that labels a queried data instance as an anomaly or nominal point. This feedback is intended to adjust the anomaly detector so that the outliers it discovers are more in tune with the expert user’s semantic understanding of the anomalies. The AAD algorithm is based on a weighted ensemble of anomaly detectors. When it receives a label from the user, it adjusts the weights on each individual ensemble member such that the anomalies rank higher in terms of their anomaly score than the outliers. The AAD approach is designed to operate in an interactive data exploration loop. In each iteration of this loop, our algorithm first selects a data instance to present to the expert as a potential anomaly and then the expert labels the instance as an anomaly or as a nominal data point. When it receives the instance label, the algorithm updates its internal model and the loop continues until a budget of B queries is spent. The goal of our approach is to maximize the total number of true anomalies in the B instances presented to the expert. We show that the AAD method performs well and in some cases doubles the number of true anomalies found compared to previous methods. In addition we present approximations that make the AAD algorithm much more computationally efficient while maintaining a desirable level of performance.

Accurate weather data is important for improving agricultural productivity in developing countries. Unfortunately, weather sensors can fail for a wide variety of reasons. One approach to detecting failed sensors is to identify statistical anomalies in the joint distribution of sensor readings. This powerful method can break down if some of the sensor readings are missing. This paper evaluates five strategies for handling missing values in anomaly detection: (a) mean imputation, (b) MAP imputation, (c) reduction (reduced-dimension anomaly detectors via feature bagging), (d) marginalization (for density estimators only), and (e) proportional distribution (for tree-based methods only). Our analysis suggests that MAP imputation and proportional distribution should give better results than mean imputation, reduction, and marginalization. These hypotheses are largely confirmed by experimental studies on synthetic data and on anomaly detection benchmark data sets using the Isolation Forest (IF), LODA, and EGMM anomaly detection algorithms. However, marginalization worked surprisingly well for EGMM, and there are exceptions where reduction works well on some benchmark problems. We recommend proportional distribution for IF, MAP imputation for LODA, and marginalization for EGMM.

Many ecological studies and conservation policies are based on field observations of species, which can be affected by systematic variability introduced by the observation process. A recently introduced causal modeling technique called 'half-sibling regression' can detect and correct for systematic errors in measurements of multiple independent random variables. However, it will remove intrinsic variability if the variables are dependent, and therefore does not apply to many situations, including modeling of species counts that are controlled by common causes. We present a technique called 'three-quarter sibling regression' to partially overcome this limitation. It can filter the effect of systematic noise when the latent variables have observed common causes. We provide theoretical justification of this approach, demonstrate its effectiveness on synthetic data, and show that it reduces systematic detection variability due to moon brightness in moth surveys.

In many applications, an anomaly detection system presents the most anomalous data instance to a human analyst, who then must determine whether the instance is truly of interest (e.g., a threat in a security setting). Unfortunately, most anomaly detectors provide no explanation about why an instance was considered anomalous, leaving the analyst with no guidance about where to begin the investigation. To address this issue, we study the problems of computing and evaluating sequential feature explanations (SFEs) for anomaly detectors. An SFE of an anomaly is a sequence of features, which are presented to the analyst one at a time (in order) until the information contained in the highlighted features is enough for the analyst to make a confident judgement about the anomaly. Since analyst effort is related to the amount of information that they consider in an investigation, an explanation’s quality is related to the number of features that must be revealed to attain confidence. In this article, we first formulate the problem of optimizing SFEs for a particular density-based anomaly detector. We then present both greedy algorithms and an optimal algorithm, based on branch-and-bound search, for optimizing SFEs. Finally, we provide a large scale quantitative evaluation of these algorithms using a novel framework for evaluating explanations. The results show that our algorithms are quite effective and that our best greedy algorithm is competitive with optimal solutions.

Exogenous state variables and rewards can slow down reinforcement learning by injecting uncontrolled variation into the reward signal. We formalize exogenous state variables and rewards and identify conditions under which an MDP with exogenous state can be decomposed into an exogenous Markov Reward Process involving only the exogenous state+reward and an endogenous Markov Decision Process defined with respect to only the endogenous rewards. We also derive a variance-covariance condition under which Monte Carlo policy evaluation on the endogenous MDP is accelerated compared to using the full MDP. Similar speedups are likely to carry over to all RL algorithms. We develop two algorithms for discovering the exogenous variables and test them on several MDPs. Results show that the algorithms are practical and can significantly speed up reinforcement learning.

Open category detection is the problem of detecting "alien" test instances that belong to categories or classes that were not present in the training data. In many applications, reliably detecting such aliens is central to ensuring the safety and accuracy of test set predictions. Unfortunately, there are no algorithms that provide theoretical guarantees on their ability to detect aliens under general assumptions. Further, while there are algorithms for open category detection, there are few empirical results that directly report alien detection rates. Thus, there are significant theoretical and empirical gaps in our understanding of open category detection. In this paper, we take a step toward addressing this gap by studying a simple, but practically-relevant variant of open category detection. In our setting, we are provided with a "clean" training set that contains only the target categories of interest and an unlabeled "contaminated” training set that contains a fraction alpha of alien examples. Under the assumption that we know an upper bound on alpha we develop an algorithm with PAC-style guarantees on the alien detection rate, while aiming to minimize false alarms. Empirical results on synthetic and standard benchmark datasets demonstrate the regimes in which the algorithm can be effective and provide a baseline for further advancements.

Anomaly detectors are often used to produce a ranked list of statistical anomalies, which are examined by human analysts in order to extract the actual anomalies of interest. This can be exceedingly difficult and time consuming when most high-ranking anomalies are false positives and not interesting from an application perspective. In this paper, we study how to reduce the analyst's effort by incorporating their feedback about whether the anomalies they investigate are of interest or not. In particular, the feedback will be used to adjust the anomaly ranking after every analyst interaction, ideally moving anomalies of interest closer to the top. Our main contribution is to formulate this problem within the framework of online convex optimization, which yields an efficient and extremely simple approach to incorporating feedback compared to the prior state-of-the-art. We instantiate this approach for the powerful class of tree-based anomaly detectors and conduct experiments on a range of benchmark datasets. The results demonstrate the utility of incorporating feedback and advantages of our approach over the state-of-the-art. In addition, we present results on a significant cybersecurity application where the goal is to detect red-team attacks in real system audit data. We show that our approach for incorporating feedback is able to significantly reduce the time required to identify malicious system entities across multiple attacks on multiple operating systems.

In many fields in computational sustainability, applications of POMDPs are inhibited by the complexity of the optimal solution. One way of delivering simple solutions is to represent the policy with a small number of alpha-vectors. We would like to find the best possible policy that can be expressed using a fixed number N of alpha-vectors. We call this the N-POMDP problem. The existing solver alpha-min approximately solves finite-horizon POMDPs with a controllable number of alpha-vectors. However alpha-min is a greedy algorithm without performance guarantees, and it is rather slow. This paper proposes three new algorithms, based on a general approach that we call alpha-min-2. These three algorithms are able to approximately solve N-POMDPs. Alpha-min-2-fast (heuristic) and alpha-min-2-p (with performance guarantees) are designed to complement an existing POMDP solver, while alpha-min-2-solve (heuristic) is a solver itself. Complexity results are provided for each of the algorithms, and they are tested on well-known benchmarks. These new algorithms will help users to interpret solutions to POMDP problems in computational sustainability.

Sample-based tree search (SBTS) is an approach to solving Markov decision problems based on constructing a lookahead search tree using random samples from a generative model of the MDP. It encompasses Monte Carlo tree search (MCTS) algorithms like UCT as well as algorithms such as sparse sampling. SBTS is well-suited to solving MDPs with large state spaces due to the relative insensitivity of SBTS algorithms to the size of the state space. The limiting factor in the performance of SBTS tends to be the exponential dependence of sample complexity on the depth of the search tree. The number of samples required to build a search tree is O((|A|B)^d), where |A| is the number of available actions, B is the number of possible random outcomes of taking an action, and d is the depth of the tree. State abstraction can be used to reduce B by aggregating random outcomes together into abstract states. Recent work has shown that abstract tree search often performs substantially better than tree search conducted in the ground state space. This paper presents a theoretical and empirical evaluation of tree search with both fixed and adaptive state abstractions. We derive a bound on regret due to state abstraction in tree search that decomposes abstraction error into three components arising from properties of the abstraction and the search algorithm. We describe versions of popular SBTS algorithms that use fixed state abstractions, and we introduce the Progressive Abstraction Refinement in Sparse Sampling (PARSS) algorithm, which adapts its abstraction during search. We evaluate PARSS as well as sparse sampling with fixed abstractions on 12 experimental problems, and find that PARSS outperforms search with a fixed abstraction and that search with even highly inaccurate fixed abstractions outperforms search without abstraction. These results establish progressive abstraction refinement as a promising basis for new tree search algorithms, and we propose directions for future work within the progressive refinement framework.

Unsupervised anomaly detection algorithms search for outliers and then predict that these outliers are the anomalies. When deployed, however, these algorithms are often criticized for high false positive and high false negative rates. One cause of poor performance is that not all outliers are anomalies and not all anomalies are outliers. In this paper, we describe an Active Anomaly Discovery (AAD) method for incorporating expert feedback to adjust the anomaly detector so that the outliers it discovers are more in tune with the expert user's semantic understanding of the anomalies. The AAD approach is designed to operate in an interactive data exploration loop. In each iteration of this loop, our algorithm first selects a data instance to present to the expert as a potential anomaly and then the expert labels the instance as an anomaly or as a nominal data point. Our algorithm updates its internal model with the instance label and the loop continues until a budget of B queries is spent. The goal of our approach is to maximize the total number of true anomalies in the B instances presented to the expert. We show that when compared to other state-of-the-art algorithms, AAD is consistently one of the best performers.

Easy-first, a search-based structured prediction approach, has been applied to many NLP tasks including dependency parsing and coreference resolution. This approach employs a learned greedy policy (action scoring function) to make easy decisions first, which constrains the remaining decisions and makes them easier. We formulate greedy policy learning in the Easy-first approach as a novel non-convex optimization problem and solve it via an efficient Majorization Minimizatoin (MM) algorithm. Results on within-document coreference and cross-document joint entity and event coreference tasks demonstrate that the proposed approach achieves statistically significant performance improvement over existing training regimes for Easy-first and is less susceptible to overfitting.

The mainstream approach to structured prediction problems in computer vision is to learn an energy function such that the solution minimizes that function. At prediction time, this approach must solve an often-challenging optimization problem. Search-based methods provide an alternative that has the potential to achieve higher performance. These methods learn to control a search procedure that constructs and evaluates candidate solutions. The recently-developed ℋC-Search method has been shown to achieve state-of-the-art results in natural language processing, but mixed success when applied to vision problems. This paper studies whether ℋC-Search can achieve similarly competitive performance on basic vision tasks such as object detection, scene labeling, and monocular depth estimation, where the leading paradigm is energy minimization. To this end, we introduce a search operator suited to the vision domain that improves a candidate solution by probabilistically sampling likely object configurations in the scene from the hierarchical Berkeley segmentation. We complement this search operator by applying the DAgger algorithm to robustly train the search heuristic so it learns from its previous mistakes. Our evaluation shows that these improvements reduce the branching factor and search depth, and thus give a significant performance boost. Our state-of-the-art results on scene labeling and depth estimation suggest that ℋC-Search provides a suitable tool for learning and inference in vision.

Tagging email is an important tactic for managing information overload. Machine learning methods can help the user with this task by predicting tags for incoming email messages. The natural user interface displays the predicted tags on the email message, and the user doesn't need to do anything unless those predictions are wrong (in which case, the user can delete the incorrect tags and add the missing tags). From a machine learning perspective, this means that the learning algorithm never receives confirmation that its predictions are correct---it only receives feedback when it makes a mistake. This can lead to slower learning, particularly when the predictions were not very confident, and hence, the learning algorithm would benefit from positive feedback. One could assume that if the user never changes any tag, then the predictions are correct, but users sometimes forget to correct the tags, presumably because they are focused on the content of the email messages and fail to notice incorrect and missing tags. The aim of this paper is to determine whether implicit feedback can provide useful additional training examples to the email prediction subsystem of TaskTracer, known as EP2 (Email Predictor 2). Our hypothesis is that the more time a user spends working on an email message, the more likely it is that the user will notice tag errors and correct them. If no corrections are made, then perhaps it is safe for the learning system to treat the predicted tags as being correct and train accordingly. This paper proposes three algorithms (and two baselines) for incorporating implicit feedback into the EP2 tag predictor. These algorithms are then evaluated using email interaction and tag correction events collected from 14 user-study participants as they performed email-directed tasks while using TaskTracer EP2. The results show that implicit feedback produces important increases in training feedback, and hence, significant reductions in subsequent prediction errors despite the fact that the implicit feedback is not perfect. We conclude that implicit feedback mechanisms can provide a useful performance boost for email tagging systems.

Researchers in AI and Operations Research employ the framework of Markov Decision Processes (MDPs) to formalize problems of sequential decision making under uncertainty. A common approach is to implement a simulator of the stochastic dynamics of the MDP and a Monte Carlo optimization algorithm that invokes this simulator to solve the MDP. The resulting software system is often realized by integrating several systems and functions that are collectively subject to failures of specification, implementation, integration, and optimization. We present these failures as queries for a computational steering visual analytic system (MDPVIS). MDPVIS addresses three visualization research gaps. First, the data acquisition gap is addressed through a general simulator-visualization interface. Second, the data analysis gap is addressed through a generalized MDP information visualization. Finally, the cognition gap is addressed by exposing model components to the user. MDPVIS generalizes a visualization for wildfire management. We use that problem to illustrate MDPVIS.

Markov Decision Process (MDP) simulators and optimization algorithms integrate several systems and functions that are collectively subject to failures of specification, implementation, integration, and optimization. We present a domain agnostic visual analytic design and implementation for testing and debugging MDPs: MDPVIS.

Monte Carlo tree search (MCTS) algorithms are a popular approach to online decision-making in Markov decision processes (MDPs). These algorithms can, however, perform poorly in MDPs with high stochastic branching factors. In this paper, we study state aggregation as a way of reducing stochastic branching in tree search. Prior work has studied formal properties of MDP state aggregation in the context of dynamic programming and reinforcement learning, but little attention has been paid to state aggregation in MCTS. Our main result is a performance loss bound for a class of value function-based state aggregation criteria in expectimax search trees. We also consider how to construct MCTS algorithms that operate in the abstract state space but require a simulator of the ground dynamics only. We find that trajectory sampling algorithms like UCT can be adapted easily, but that sparse sampling algorithms present difficulties. As a proof of concept, we experimentally confirm that state aggregation can improve the finite-sample performance of UCT.

Scripts have been proposed to model the stereotypical event sequences found in narratives. They can be applied to make a variety of inferences including fillinggaps in the narratives and resolving ambiguous references. This paper proposes the first formal frameworkfor scripts based on Hidden Markov Models (HMMs). Our framework supports robust inference and learning algorithms, which are lacking in previous clustering models. We develop an algorithm for structure andparameter learning based on Expectation Maximizationand evaluate it on a number of natural datasets. The results show that our algorithm is superior to several informed baselines for predicting missing events in partialobservation sequences.

Chao Ma, Janardhan Rao Doppa, J. Walker Orr, Prashanth Mannem, Xiaoli Fern, Tom Dietterich, Prasad Tadepalli. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014.

In the Superset Label Learning (SLL) problem, weak supervision is provided in the form of a \it superset of labels that contains the true label. If the classifier predicts a label outside of the superset, it commits a \it superset error. Most existing SLL algorithms learn a multiclass classifier by minimizing the superset error. However, only limited theoretical analysis has been dedicated to this approach. In this paper, we analyze Empirical Risk Minimizing learners that use the superset error as the empirical risk measure. SLL data can arise either in the form of independent instances or as multiple-instance bags. For both scenarios, we give the conditions for ERM learnability and sample complexity for the realizable case.

The Collective Graphical Model (CGM) models a population of independent and identically distributed individuals when only collective statistics (i.e., counts of individuals) are observed. Exact inference in CGMs is intractable, and previous work has explored Markov Chain Monte Carlo (MCMC) and MAP approximations for learning and inference. This paper studies Gaussian approximations to the CGM. As the population grows large, we show that the CGM distribution converges to a multivariate Gaussian distribution (GCGM) that maintains the conditional independence properties of the original CGM. If the observations are exact marginals of the CGM or marginals that are corrupted by Gaussian noise, inference in the GCGM approximation can be computed efficiently in closed form. If the observations follow a different noise model (e.g., Poisson), then expectation propagation provides efficient and accurate approximate inference. The accuracy and speed of GCGM inference is compared to the MCMC and MAP methods on a simulated bird migration problem. The GCGM matches or exceeds the accuracy of the MAP method while being significantly faster.

Often the most practical way to define a Markov Decision Process (MDP) is as a simulator that, given a state and an action, produces a resulting state and immediate reward sampled from the corresponding distributions. Simulators in natural resource management can be very expensive to execute, so that the time required to solve such MDPs is dominated by the number of calls to the simulator. This paper presents an algorithm, DDV, that combines improved confidence intervals on the Q values (as in interval estimation) with a novel upper bound on the discounted state occupancy probabilities to intelligently choose state-action pairs to explore. We prove that this algorithm terminates with a policy whose value is within epsilon of the optimal policy (with probability 1-delta) after making only polynomially-many calls to the simulator. Experiments on one benchmark MDP and on an MDP for invasive species management show very large reductions in the number of simulator calls required.

Archived data from the WSR-88D network of weather radars in the US hold detailed information about the continent-scale migratory movements of birds over the last 20 years. However, significant technical challenges must be overcome to understand this information and harness its potential for science and conservation. We present an approximate Bayesian inference algorithm to reconstruct the velocity fields of birds migrating in the vicinity of a radar station. This is part of a larger project to quantify bird migration at large scales using weather radar data.

In the era of large scientific data sets, there is an urgent need for methods to automatically prioritize data for review. At the same time, for any automated method to be adopted by scientists, it must make decisions that they can understand and trust. In this paper, we propose Discovery through Eigenbasis Modeling of Uninteresting Data (DEMUD), which uses principal components modeling and reconstruction error to prioritize data. DEMUD’s major advance is to offer domain-specific explanations for its prioritizations. We evaluated DEMUD’s ability to quickly identify diverse items of interest and the value of the explanations it provides. We found that DEMUD performs as well or better than existing class discovery methods and provides, uniquely, the first explanations for why those items are of interest. Further, in collaborations with planetary scientists, we found that DEMUD (1) quickly identifies very rare items of scientific value, (2) maintains high diversity in its selections, and (3) provides explanations that greatly improve human classification accuracy.

We study the problem of approximate inference in collective graphical models (CGMs), which were recently introduced to model the problem of learning and inference with noisy aggregate observations. We first analyze the complexity of inference in CGMs: unlike inference in conventional graphical models, exact inference in CGMs is NP-hard even for tree-structured models. We then develop a tractable convex approximation to the NP-hard MAP inference problem in CGMs, and show how to use MAP inference for approximate marginal inference within the EM framework. We demonstrate empirically that these approximation techniques can reduce the computational cost of inference by two orders of magnitude and the cost of learning by at least an order of magnitude while providing solutions of equal or better quality.

This paper reports on methods and results of an applied research project by a team consisting of SAIC and four universities to develop, integrate, and evaluate new approaches to detect the weak signals characteristic of insider threats on organizations' information systems. Our system combines structural and semantic information from a real corporate database of monitored activity on their users' computers to detect independently developed red team inserts of malicious insider activities. We have developed and applied multiple algorithms for anomaly detection based on suspected scenarios of malicious insider behavior, indicators of unusual activities, high-dimensional statistical patterns, temporal sequences, and normal graph evolution. Algorithms and representations for dynamic graph processing provide the ability to scale as needed for enterprise-level deployments on real-time data streams. We have also developed a visual language for specifying combinations of features, baselines, peer groups, time periods, and algorithms to detect anomalies suggestive of instances of insider threat behavior. We defined over 100 data features in seven categories based on approximately 5.5 million actions per day from approximately 5,500 users. We have achieved area under the ROC curve values of up to 0.979 and lift values of 65 on the top 50 user-days identified on two months of real data.

In the superset label learning problem (SLL), each training instance provides a set of candidate labels of which one is the true label of the instance. As in ordinary regression, the candidate label set is a noisy version of the true label. In this work, we solve the problem by maximizing the likelihood of the candidate label sets of training instances. We propose a probabilistic model, the Logistic Stick- Breaking Conditional Multinomial Model (LSB-CMM), to do the job. The LSB- CMM is derived from the logistic stick-breaking process. It first maps data points to mixture components and then assigns to each mixture component a label drawn from a component-specific multinomial distribution. The mixture components can capture underlying structure in the data, which is very useful when the model is weakly supervised. This advantage comes at little cost, since the model introduces few additional parameters. Experimental tests on several real-world problems with superset labels show results that are competitive or superior to the state of the art. The discovered underlying structures also provide improved explanations of the classification predictions.

We present a novel ensemble architecture for learning problem-solving techniques from a very small number of expert solutions and demonstrate its effectiveness in a complex real-world domain. The key feature of our “Generalized Integrated Learning Architecture” (GILA) is a set of heterogeneous independent learning and reasoning (ILR) components, coordinated by a central meta-reasoning executive (MRE). The ILRs are weakly coupled in the sense that all coordination during learning and performance happens through the MRE. Each ILR learns independently from a small number of expert demonstrations of a complex task. During performance, each ILR proposes partial solutions to subproblems posed by the MRE, which are then selected from and pieced together by the MRE to produce a complete solution. The heterogeneity of the learner-reasoners allows both learning and problem solving to be more effective because their abilities and biases are complementary and synergistic. We describe the application of this novel learning and problem solving architecture to the domain of airspace management, where multiple requests for the use of airspaces need to be deconflicted, reconciled, and managed automatically. Formal evaluations show that our system performs as well as or better than humans after learning from the same training data. Furthermore, GILA outperforms any individual ILR run in isolation, thus demonstrating the power of the ensemble architecture for learning and problem solving.

Important ecological phenomena are often observed indirectly. Consequently, probabilistic latent variable models provide an important tool, because they can include explicit models of the ecological phenomenon of interest and the process by which it is observed. However, existing latent variable methods rely on hand-formulated parametric models, which are expensive to design and require extensive preprocessing of the data. Nonparametric methods (such as regression trees) automate these decisions and produce highly accurate models. However, existing tree methods learn direct mappings from inputs to outputs — they cannot be applied to latent variable models. This paper describes a methodology for integrating nonparametric tree methods into probabilistic latent variable models by extending functional gradient boosting. The approach is presented in the context of occupancy-detection (OD) modeling, where the goal is to model the distribution of a species from imperfect detections. Experiments on 12 real and 3 synthetic bird species compare standard and tree-boosted OD models (latent variable models) with standard and tree-boosted logistic regression models (without latent structure). All methods perform similarly when predicting the observed variables, but the OD models learn better representations of the latent process. Most importantly, tree-boosted OD models learn the best latent representations when nonlinearities and interactions are present.

There are many settings in which we wish to fit a model of the behavior of individuals but where our data consist only of aggregate information (counts or low-dimensional contingency tables). This paper introduces Collective Graphical Models---a framework for modeling and probabilistic inference that operates directly on the sufficient statistics of the individual model. We derive a highly-efficient Gibbs sampling algorithm for sampling from the posterior distribution of the sufficient statistics conditioned on noisy aggregate observations, prove its correctness, and demonstrate its effectiveness experimentally.

We consider the problem of learning rules from natural language text sources. These sources, such as news articles and web texts, are created by a writer to communicate information to a reader, where the writer and reader share substantial domain knowledge. Consequently, the texts tend to be concise and mention the minimum information necessary for the reader to draw the correct conclusions. We study the problem of learning domain knowledge from such concise texts, which is an instance of the general problem of learning in the presence of missing data. However, unlike standard approaches to missing data, in this setting we know that facts are more likely to be missing from the text in cases where the reader can infer them from the facts that are mentioned combined with the domain knowledge. Hence, we can explicitly model this "missingness" process and invert it via probabilistic inference to learn the underlying domain knowledge. This paper introduces a mention model that models the probability of facts being mentioned in the text based on what other facts have already been mentioned and domain knowledge in the form of Horn clause rules. Learning must simultaneously search the space of rules and learn the parameters of the mention model. We accomplish this via an application of Expectation Maximization within a Markov Logic framework. An experimental evaluation on synthetic and natural text data shows that the method can learn accurate rules and apply them to new texts to make correct inferences. Experiments also show that the method out-performs the standard EM approach that assumes mentions are missing at random.

The combination of local features, complementary feature types, and relative position information has been successfully applied to many object-class recognition tasks. Stacking is a common classification approach that combines the results from multiple classifiers, having the added benefit of allowing each classifier to handle a different feature space. However, the standard stacking method by its own nature discards any spatial information contained in the features, because only the combination of raw classification scores are input to the final classifier. The object-class recognition method proposed in this paper combines different feature types in a new stacking framework that efficiently quantizes input data and boosts classification accuracy, while allowing the use of spatial information. This classification method is applied to the task of automated insect-species identification for biomonitoring purposes. The test data set for this work contains 4722 images with 29 insect species, belonging to the three most common orders used to measure stream water quality, several of which are closely related and very difficult to distinguish. The specimens are in different 3D positions, different orientations, and different developmental and degradation stages with wide intra-class variation. On this very challenging data set, our new algorithm outperforms other classifiers, showing the benefits of using spatial information in the stacking framework with multiple dissimilar feature types.

Helping computer users rapidly locate files in their folder hierarchies is a practical research problem involving both intelligent systems and user interface design. This article reports on FolderPredictor, a software system that can reduce the cost of locating files in hierarchical folders. FolderPredictor applies a cost-sensitive prediction algorithm to the user's previous file access information to predict the next folder that will be accessed. Experimental results show that, on average, FolderPredictor reduces the number of clicks spent on locating a file by 50%. Several variations of the cost-sensitive prediction algorithm are discussed. An experimental study shows that the best algorithm among them is a mixture of the most recently used (MRU) folder and the cost-sensitive predictions. Furthermore, FolderPredictor does not require users to adapt to a new interface, but rather meshes with the existing interface for opening files on the Windows platform.

The ecological sciences have benefited greatly from recent advances in wireless sensor technologies. These technologies allow researchers to deploy networks of automated sensors, which can monitor a landscape at very fine temporal and spatial scales. However, these networks are subject to harsh conditions, which lead to malfunctions in individual sensors and failures in network communications. The resulting data streams often exhibit incorrect data measurements and missing values. Identifying and correcting these is time-consuming and error-prone. We present a method for real-time automated data quality control (QC) that exploits the spatial and temporal correlations in the data to distinguish sensor failures from valid observations. The model adapts to each deployment site by learning a Bayesian network structure that captures spatial relationships between sensors, and it extends the structure to a dynamic Bayesian network to incorporate temporal correlations. This model is able to flag faulty observations and predict the true values of the missing or corrupt readings. The performance of the model is evaluated on data collected by the SensorScope Project. The results show that the spatiotemporal model demonstrates clear advantages over models that include only temporal or only spatial correlations, and that the model is capable of accurately imputing corrupted values.

We consider the problem of incorporating end-user advice into reinforcement learning (RL). In our setting, the learner alternates between practicing, where learning is based on actual world experience, and end-user critique sessions where advice is gathered. During each critique session the end-user is allowed to analyze a trajectory of the current policy and then label an arbitrary subset of the available actions as good or bad. Our main contribution is an approach for integrating all of the information gathered during practice and critiques in order to effectively optimize a parametric policy. The approach optimizes a loss function that linearly combines losses measured against the world experience and the critique data. We evaluate our approach using a prototype system for teaching tactical battle behavior in a real-time strategy game engine. Results are given for a significant evaluation involving ten end-users showing the promise of this approach and also highlighting challenges involved in inserting end-users into the RL loop.

In the field of Human-Computer Interaction, provenance refers to the history and genealogy of a document or file. Provenance helps us to understand the evolution and relationships of files; how and when different versions of a document were created, or how different documents in a collection build on each other through copy-paste events. Though methods for tracking provenance and the subsequent use of this meta-data have been proposed and developed into tools, there have been no studies documenting the types and frequency of provenance events in typical computer use. This is knowledge essential for the design of efficient query methods and information displays. We conducted a longitudinal study of knowledge workers at Intel Corporation tracking provenance events in their computer use. We also interviewed knowledge workers to determine the effectiveness of provenance cues for document recall. Our data shows that provenance relationships are common, and provenance cues aid recall.

This paper proposes an image classification method based on extracting image features using Haar random forests and combining them with a spatial matching kernel SVM. The method works by combining multiple efficient, yet powerful, learning algorithms at every stage of the recognition process. On the task of identifying aquatic stonefly larvae, the method has state-of-the-art or better performance, but with much higher efficiency.

Current work in object categorization discriminates among objects that typically possess gross differences which are readily apparent. However, many applications require making much finer distinctions. We address an insect categorization problem that is so challenging that even trained human experts cannot readily categorize images of insects considered in this paper. The state of the art that uses visual dictionaries, when applied to this problem, yields mediocre results (16.1% error). Three possible explanations for this are (a) the dictionaries are unsupervised, (b) the dictionaries lose the detailed information contained in each keypoint, and (c) these methods rely on hand-engineered decisions about dictionary size. This paper presents a novel, dictionary-free methodology. A random forest of trees is first trained to predict the class of an image based on individual keypoint descriptors. A unique aspect of these trees is that they do not make decisions but instead merely record evidence-i.e., the number of descriptors from training examples of each category that reached each leaf of the tree. We provide a mathematical model showing that voting evidence is better than voting decisions. To categorize a new image, descriptors for all detected keypoints are “dropped” through the trees, and the evidence at each leaf is summed to obtain an overall evidence vector. This is then sent to a second-level classifier to make the categorization decision. We achieve excellent performance (6.4% error) on the 9-class STONEFLY9 data set. Also, our method achieves an average AUC of 0.921 on the PASCAL06 VOC, which places it fifth out of 21 methods reported in the literature and demonstrates that the method also works well for generic object categorization.

Codebook-based representations are widely employed in the classification of complex objects such as images and documents. Most previous codebook-based methods construct a single codebook via clustering that maps a bag of low-level features into a fixed-length histogram that describes the distribution of these features. This paper describes a simple yet effective framework for learning multiple non-redundant codebooks that produces surprisingly good results. In this framework, each codebook is learned in sequence to extract discriminative information that was not captured by preceding codebooks and their corresponding classifiers. We apply this framework to two application domains: visual object categorization and document classification. Experiments on large classification tasks show substantial improvements in performance compared to a single codebook or codebooks learned in a bagging style.

Intelligent desktop assistants could provide more help for users if they could learn models of the users' workflows. However, discovering desktop workflows is difficult because they unfold over extended periods of time (days or weeks) and they are interleaved with many other workflows because of user multi-tasking. This paper describes an approach to discovering desktop workflows based on rich instrumentation of information flow actions such as copy/paste, SaveAs, file copy, attach file to email message, and save attachment. These actions allow us to construct a graph whose nodes are files, email messages, and web pages and whose edges are these information flow actions. A class of workflows that we call work procedures can be discovered by applying graph mining algorithms to find frequent subgraphs. This paper describes an algorithm for mining frequent closed connected subgraphs and then describes the results of applying this method to data collected from a group of real users.

The TaskTracer system allows knowledge workers to define a set of activities that characterize their desktop work. It then associates with each user-defined activity the set of resources that the user accesses when performing that activity. In order to correctly associate resources with activities and provide useful activity-related services to the user, the system needs to know the current activity of the user at all times. It is often convenient for the user to explicitly declare which activity he/she is working on. But frequently the user forgets to do this. TaskTracer applies machine learning methods to detect undeclared activity switches and predict the correct activity of the user. This paper presents TaskPredictor2, a complete redesign of the activity predictor in TaskTracer and its notification user interface. TaskPredictor2 applies a novel online learning algorithm that is able to incorporate a richer set of features than our previous predictors. We prove an error bound for the algorithm and present experimental results that show improved accuracy and a 180-fold speedup on real user data. The user interface supports negotiated interruption and makes it easy for the user to correct both the predicted time of the task switch and the predicted activity.

Real-time problems, in which the learning must be fast and the importance of the features might be changing, pose a challenge to machine learning algorithms. To learn robust classifiers in such nonstationary environments, it is essential not to assign too much weight to any single feature. We solve the problems by combining regularization mechanisms with online large margin learning algorithms. We prove bounds on their error and show that removing features with small weights has little influence on the accuracy, suggesting that these methods exhibit feature selection ability. We show that such regularized learning algorithms automatically decrease the influence of the old training instances and focus on the more recent ones. This makes them especially attractive in the dynamic environments. We present experimental results on real datasets to show the merit of our algorithms.

We present an algorithm, HI-MAT (Hierarchy Induction via Models And Trajectories), that discovers MAXQ task hierarchies by applying dynamic Bayesian network models to a successful trajectory from a source reinforcement learning task. HI-MAT discovers subtasks by analyzing the causal and temporal relationships among the actions in the trajectory. Under appropriate assumptions, HI-MAT induces hierarchies that are consistent with the observed trajectory and have compact value-function tables employing safe state abstractions. We demonstrate empirically that HI-MAT constructs compact hierarchies that are comparable to manually-engineered hierarchies and facilitate significant speedup in learning when transferred to a target task.

Visual dictionaries are widely employed in object recognition to map unordered bags of local region descriptors into feature vectors for image classification. Most visual dictionaries have been constructed by unsupervised clustering. This paper presents an efficient discriminative approach, called Iterative Discriminative Clustering (IDC), for dictionary learning. In this approach, each dictionary entry is defined by a representative value and a learned distance metric. In IDC algorithm, the dictionary entries are initialized by unsupervised clustering and then locally adapted to improve their discriminative power. Motivated by studies of the characteristics of individual dictionary entries, we employ bagged decision lists (BDL) as our image classifier in order to explore the conjunctions of small number of informative dictionary entries for classification. Experiments on benchmark object recognition datasets show that the system based on the new discriminative dictionaries and BDL classifier give performance comparable or superior to the state-of-art generic object recognition approaches.

This paper presents a new structure-based interest region detector called principal curvature-based regions (PCBR) which we use for object class recognition. The PCBR interest operator detects stable watershed regions within the multi-scale principal curvature image. To detect robust watershed regions, we "clean" a principal curvature image by combining a grayscale morphological close with our new "eigenvectorflow" hysteresis threshold. Robustness across scales is achieved by selecting the maximally stable regions across consecutive scales. PCBR typically detects distinctive patterns distributed evenly on the objects and it shows significant robustness to local intensity perturbations and intra-class variations. We evaluate PCBR both qualitatively (through visual inspection) and quantitatively (by measuring repeatability and classification accuracy in real-world object-class recognition problems). Experiments on different benchmark datasets show that PCBR is comparable or superior to state-of-art detectors for both feature matching and object recognition. Moreover, we demonstrate the application of PCBR to symmetry detection.

Intelligent desktop environments allow the desktop user to define a set of projects or activities that characterize the user's desktop work. These environments then attempt to identify the current activity of the user in order to provide various kinds of assistance. These systems take a hybrid approach in which they allow the user to declare their current activity but they also employ learned classifiers to predict the current activity to cover those cases where the user forgets to declare the current activity. The classifiers must be trained on the very noisy data obtained from the user's activity declarations. Instead of asking the user to review and relabel the data manually, we employ an active EM algorithm that combines the EM algorithm and active learning. EM can be viewed as retraining on its own predictions. To make it more robust, we only retrain on those predictions that are made with high confidence. For active learning, we make a small number of queries to the user based on the most uncertain instances. Experimental results on real users show this active EM algorithm can significantly improve the prediction precision, and that it performs better than either EM or active learning alone.

There has been little research into how end users might be able to communicate advice to machine learning systems. If this resource--the users themselves--could somehow work hand-in-hand with machine learning systems, the accuracy of learning systems could be improved and the users' understanding and trust of the system could improve as well. We conducted a think-aloud study to see how willing users were to provide feedback and to understand what kinds of feedback users could give. Users were shown explanations of machine learning predictions and asked to provide feedback to improve the predictions. We found that users had no difficulty providing generous amounts of feedback. The kinds of feedback ranged from suggestions for reweighting of features to proposals for new features, feature combinations, relational features, and wholesale changes to the learning algorithm. The results show that user feedback has the potential to significantly improve machine learning systems, but that learning algorithms need to be extended in several ways to be able to assimilate this feedback.

This paper describes a fully automated stone fly-larvae classification system using a local features approach. It compares the three region detectors employed by the system: the Hessian-affine detector, the Kadir entropy detector and a new detector we have developed called the principal curvature based region detector (PCBR). It introduces a concatenated feature histogram (CFH) methodology that uses histograms of local region descriptors as feature vectors for classification and compares the results using this methodology to that of Opelt [Opelt, A, et.al., 2006.] on three stonefly identification tasks. Our results indicate that the PCBR detector outperforms the other two detectors on the most difficult discrimination task and that the use of all three detectors outperforms any other configuration. The CFH methodology also outperforms the Opelt methodology in these tasks

Helping computer users rapidly locate files in their folder hierarchies has become an important research topic in today's intelligent user interface design. This paper reports on FolderPredictor, a software system that can reduce the cost of locating files in hierarchical folders. FolderPredictor applies a cost-sensitive prediction algorithm to the user's previous file access information to predict the next folder that will be accessed. Experimental results show that, on average, FolderPredictor reduces the cost of locating a file by 50%. Another advantage of FolderPredictor is that it does not require users to adapt to a new interface, but rather meshes with the existing interface for opening files on the Windows platform.

The TaskTracer system seeks to help multi-tasking users manage the resources that they create and access while carrying out their work activities. It does this by associating with each user-defined activity the set of files, folders, email messages, contacts, and web pages that the user accesses when performing that activity. The initial TaskTracer system relies on the user to notify the system each time the user changes activities. However, this is burdensome, and users often forget to tell TaskTracer what activity they are working on. This paper introduces TaskPredictor, a machine learning system that attempts to predict the user’s current activity. TaskPredictor has two components: one for general desktop activity and another specifically for email. TaskPredictor achieves high prediction precision by combining three techniques: (a) feature selection via mutual information, (b) classification based on a confidence threshold, and (c) a hybrid design in which a Naive Bayes classifier estimates the classification confidence but where the actual classification decision is made by a support vector machine. This paper provides experimental results on data collected from Task-Tracer users.

First-order probabilistic models allow us to model situations in which a random variable in the first-order model may have a large and varying numbers of parent variables in the ground ("unrolled") model. One approach to compactly describing such models is to independently specify the probability of a random variable conditioned on each individual parent (or small sets of parents) and then combine these conditional distributions via a combining rule (e.g., Noisy-OR). This paper presents algorithms for learning with combining rules. Specifically, algorithms based on gradient descent and expectation maximization are derived, implemented, and evaluated on synthetic data and on a real-world task. The results demonstrate that the algorithms are able to learn the parameters of both the individual parent-target distributions and the combining rules.

This paper reports on TaskTracer --- a software system being designed to help highly multitasking knowledge workers rapidly locate, discover, and reuse past processes they used to successfully complete tasks. The system monitors users' interaction with a computer, collects detailed records of users' activities and resources accessed, associates (automatically or with users' assistance) each interaction event with a particular task, enables users to access records of past activities and quickly restore task contexts. We present a novel Publisher-Subscriber architecture for collecting and processing users' activity data, describe several different user interfaces tried with TaskTracer, and discuss the possibility of applying machine learning techniques to recognize/predict users' tasks.

This paper studies the problem of learning diagnostic policies from training examples. A diagnostic policy is a complete description of the decision-making actions of a diagnostician (i.e., tests followed by a diagnostic decision) for all possible combinations of test results. An optimal diagnostic policy is one that minimizes the expected total cost, which is the sum of measurement costs and misdiagnosis costs. In most diagnostic settings, there is a tradeoff between these two kinds of costs. This paper formalizes diagnostic decision making as a Markov Decision Process (MDP). The paper introduces a new family of systematic search algorithms based on the AO* algorithm to solve this MDP. To make AO* efficient, the paper describes an admissible heuristic that enables AO* to prune large parts of the search space. The paper also introduces several greedy algorithms including some improvements over previously-published methods. The paper then addresses the question of learning diagnostic policies from examples. When the probabilities of diseases and test results are computed from training data, there is a great danger of overfitting. To reduce overfitting, regularizers are integrated into the search algorithms. Finally, the paper compares the proposed methods on five benchmark diagnostic data sets. The studies show that in most cases the systematic search methods produce better diagnostic policies than the greedy methods. In addition, the studies show that for training sets of realistic size, the systematic search algorithms are practical on today's desktop computers.

Conditional Random Fields (CRFs; Lafferty, McCallum, & Pereira, 2001) provide a flexible and powerful model for learning to assign labels to elements of sequences in such applications as part-of-speech tagging, text-to-speech mapping, protein and DNA sequence analysis, and information extraction from web pages. However, existing learning algorithms are slow, particularly in problems with large numbers of potential input features. This paper describes a new method for training CRFs by applying Friedman's (1999) gradient tree boosting method. In tree boosting, the CRF potential functions are represented as weighted sums of regression trees. Regression trees are learned by stage-wise optimizations similar to Adaboost, but with the objective of maximizing the conditional likelihood P(Y|X) of the CRF model. By growing regression trees, interactions among features are introduced only as needed, so although the parameter space is potentially immense, the search algorithm does not explicitly consider the large space. As a result, gradient tree boosting scales linearly in the order of the Markov model and in the order of the feature interactions, rather than exponentially like previous algorithms based on iterative scaling and gradient descent.

The standard model of supervised learning assumes that training and test data are drawn from the same underlying distribution. This paper explores an application in which a second, auxiliary, source of data is available drawn from a different distribution. This auxiliary data is more plentiful, but of significantly lower quality, than the training and test data. In the SVM framework, a training example has two roles: (a) as a data point to constrain the learning process and (b) as a candidate support vector that can form part of the definition of the classifier. The paper considers using the auxiliary data in either (or both) of these roles. This auxiliary data framework is applied to a problem of classifying images of leaves of maple and oak trees using a kernel derived from the shapes of the leaves. Experiments show that when the training data set is very small, training with auxiliary data can produce large improvements in accuracy, even when the auxiliary data is significantly different from the training (and test) data. The paper also introduces techniques for adjusting the kernel scores of the auxiliary data points to make them more comparable to the training data points.

We present three ways of combining linear programming with the kernel trick to find value function approximations for reinforcement learning. One formulation is based on SVM regression; the second is based on the Bellman equation; and the third seeks only to ensure that good moves have an advantage over bad moves. All formu(cid:173) lations attempt to minimize the number of support vectors while fitting the data. Experiments in a difficult, synthetic maze problem show that all three formulations give excellent performance, but the advantage formulation is much easier to train. Unlike policy gradi(cid:173) ent methods, the kernel methods described here can easily 'adjust the complexity of the function approximator to fit the complexity of the value function.

We address the problem of non-convergence of online reinforcement learning algorithms (e.g., Q learning and SARSA(A)) by adopt(cid:173) ing an incremental-batch approach that separates the exploration process from the function fitting process. Our BFBP (Batch Fit to Best Paths) algorithm alternates between an exploration phase (during which trajectories are generated to try to find fragments of the optimal policy) and a function fitting phase (during which a function approximator is fit to the best known paths from start states to terminal states). An advantage of this approach is that batch value-function fitting is a global process, which allows it to address the tradeoffs in function approximation that cannot be handled by local, online algorithms. This approach was pioneered by Boyan and Moore with their GROWSUPPORT and ROUT al(cid:173) gorithms. We show how to improve upon their work by applying a better exploration process and by enriching the function fitting procedure to incorporate Bellman error and advantage error mea(cid:173) sures into the objective function. The results show improved per(cid:173) formance on several benchmark problems.

We describe an application of data mining and decision analysis to the problem of die-level functional test in integrated circuit manufacturing.Integrated circuits are fabricated on large wafers that can hold hundreds of individual chips ("die").In current practice, large and expensive machines test each of these die to check that they are functioning properly (die-level functional test; DLFT), and then the wafers are cut up, and the good die are assembled into packages and connected to the package pins.Finally, the resulting packages are tested to ensure that the final product is functioning correctly.The purpose of die-level functional test is to avoid the expense of packaging bad die and to provide rapid feedback to the fabrication process by detecting die failures.The challenge for a decision-theoretic approach is to reduce the amount of DLFT (and the associated costs) while still providing process feedback.We describe a decision-theoretic approach to DLFT in which historical test data is mined to create a probabilistic model of patterns of die failure.This model is combined with greedy value-of-information computations to decide in real time which die to test next and when to stop testing.We report the results of several experiments that demonstrate the ability of this procedure to make good testing decisions, good stopping decisions, and to detect anomalous die.Based on experiments with historical test data from Hewlett Packard Company, the resulting system has the potential to improve profits on mature IC products.

This paper presents a new approach to hierarchical reinforcement learning based on decomposing the target Markov decision process (MDP) into a hierarchy of smaller MDPs and decomposing the value function of the target MDP into an additive combination of the value functions of the smaller MDPs. The decomposition, known as the MAXQ decomposition, has both a procedural semantics---as a subroutine hierarchy---and a declarative semantics---as a representation of the value function of a hierarchical policy. MAXQ unifies and extends previous work on hierarchical reinforcement learning by Singh, Kaelbling, and Dayan and Hinton. It is based on the assumption that the programmer can identify useful subgoals and define subtasks that achieve these subgoals. By defining such subgoals, the programmer constrains the set of policies that need to be considered during reinforcement learning. The MAXQ value function decomposition can represent the value function of any policy that is consistent with the given hierarchy. The decomposition also creates opportunities to exploit state abstractions, so that individual MDPs within the hierarchy can ignore large parts of the state space. This is important for the practical application of the method. This paper defines the MAXQ hierarchy, proves formal results on its representational power, and establishes five conditions for the safe use of state abstractions. The paper presents an online model-free learning algorithm, MAXQ-Q, and proves that it converges with probability 1 to a kind of locally-optimal policy known as a recursively optimal policy, even in the presence of the five kinds of state abstraction. The paper evaluates the MAXQ representation and MAXQ-Q through a series of experiments in three domains and shows experimentally that MAXQ-Q (with state abstractions) converges to a recursively optimal policy much faster than flat Q learning. The fact that MAXQ learns a representation of the value function has an important benefit: it makes it possible to compute and execute an improved, non-hierarchical policy via a procedure similar to the policy improvement step of policy iteration. The paper demonstrates the effectiveness of this non-hierarchical execution experimentally. Finally, the paper concludes with a comparison to related work and a discussion of the design tradeoffs in hierarchical reinforcement learning.

This article reviews five approximate statistical tests for determining whether one learning algorithm outperforms another on a particular learning task. These tests are compared experimentally to determine their probability of incorrectly detecting a difference when no difference exists (type I error). Two widely used statistical tests are shown to have high probability of type I error in certain situations and should never be used: a test for difference of two proportions and a paired-differences t test based on taking several random train-test splits. A third test, a paired-differences t test based on 10-fold cross-validation, exhibits somewhat elevated probability of type I error. A fourth test, McNemar's test, is shown to have low type I error. The fifth test is a new test, 5 x 2 cv, based on five iterations of twofold cross-validation. Experiments show that this test also has acceptable type I error. The article also measures the power (ability to detect algorithm differences when they do exist) of these tests. The cross-validated t test is the most powerful. The 5 x 2 cv test is shown to be slightly more powerful than McNemar's test. The choice of the best test is determined by the computational cost of running the learning algorithm. For algorithms that can be executed only once, McNemar's test is the only test with acceptable type I error. For algorithms that can be executed 10 times, the 5 x 2 cv test is recommended, because it is slightly more powerful and because it directly measures variation due to the choice of training set.

Multiclass learning problems involve finding a definitionfor an unknown function f(x) whose range is a discrete setcontaining k > 2 values (i.e., k ``classes''). Thedefinition is acquired by studying collections of training examples ofthe form [x_i, f (x_i)]. Existing approaches tomulticlass learning problems include direct application of multiclassalgorithms such as the decision-tree algorithms C4.5 and CART,application of binary concept learning algorithms to learn individualbinary functions for each of the k classes, and application ofbinary concept learning algorithms with distributed outputrepresentations. This paper compares these three approaches to a newtechnique in which error-correcting codes are employed as adistributed output representation. We show that these outputrepresentations improve the generalization performance of both C4.5and backpropagation on a wide range of multiclass learning tasks. Wealso demonstrate that this approach is robust with respect to changesin the size of the training sample, the assignment of distributedrepresentations to particular classes, and the application ofoverfitting avoidance techniques such as decision-tree pruning.Finally, we show that---like the other methods---the error-correctingcode technique can provide reliable class probability estimates.Taken together, these results demonstrate that error-correcting outputcodes provide a general-purpose method for improving the performanceof inductive learning programs on multiclass problems.