Dental disease is a prevalent chronic condition associated with substantial financial burden, personal suffering, and increased risk of systemic diseases. Despite widespread recommendations for twice-daily tooth brushing, adherence to recommended oral self-care behaviors remains sub-optimal due to factors such as forgetfulness and disengagement. To address this, we developed Oralytics, a mHealth intervention system designed to complement clinician-delivered preventative care for marginalized individuals at risk for dental disease. Oralytics incorporates an online reinforcement learning algorithm to determine optimal times to deliver intervention prompts that encourage oral self-care behaviors. We have deployed Oralytics in a registered clinical trial. The deployment required careful design to manage challenges specific to the clinical trials setting in the U.S. In this paper, we (1) highlight key design decisions of the RL algorithm that address these challenges and (2) conduct a re-sampling analysis to evaluate algorithm design decisions. A second phase (randomized control trial) of Oralytics is planned to start in spring 2025.

Within batch reinforcement learning, safe policy improvement seeks to ensure that the learned policy performs at least as well as the behavior policy that generated the dataset. The core challenge is seeking improvements while balancing risk when many state-action pairs may be infrequently visited. In this work, we introduce Decision Points RL (DPRL), an algorithm that restricts the set of state-action pairs (or regions for continuous states) considered for improvement. DPRL ensures high-confidence improvement in densely visited states (called ‘decision points’) while still utilizing data from sparsely visited states by using them for trajectory-based value estimates. By selectively limiting the state-actions where the policy deviates from the behavior, we achieve tighter theoretical guarantees that depend only on the counts of frequently observed state-action pairs rather than on state-action space size. Our empirical results confirm DPRL provides both safety and performance improvements across synthetic and real-world applications.

We provide new connections between two distinct federated learning approaches based on (i) ADMM and (ii) Variational Bayes (VB), and propose new variants by combining their complementary strengths. Specifically, we show that the dual variables in ADMM naturally emerge through the "site" parameters used in VB with isotropic Gaussian covariances. Using this, we derive two versions of ADMM from VB that use flexible covariances and functional regularisation, respectively. Through numerical experiments, we validate the improvements obtained in performance. The work shows connection between two fields that are believed to be fundamentally different and combines them to improve federated learning.

When explaining machine learning models, it is important for explanations to have certain properties like faithfulness, robustness, smoothness, low complexity, etc. However, many properties are in tension with each other, making it challenging to achieve them simultaneously. For example, reducing the complexity of an explanation can make it less expressive, compromising its faithfulness. The ideal balance of trade-offs between properties tends to vary across different tasks and users. Motivated by these varying needs, we aim to find explanations that make optimal trade-offs while allowing for transparent control over the balance between different properties. Unlike existing methods that encourage desirable properties implicitly through their design, our approach optimizes explanations explicitly for a linear mixture of multiple properties. By adjusting the mixture weights, users can control the balance between those properties and create explanations with precisely what is needed for their particular task.

The Fugl Meyer Assessment (FMA) is a widely-used assessment for tracking motor function recovery post-stroke. Due to the limited access to rehabilitation, there exists a need for remote and automated assessment solutions. Wearable sensors and data-driven methods have shown promise for enabling automatic upper extremity FMA (FMA-UE) estimation, but minimizing user input motion and aligning with current clinical activities will aid the adoption of sensor-based assessments. In this work, we present an FMA-UE estimator which can make score predictions for a key subset of the assessment (70$\% $ of all items) using data from inertial measurement units (IMUs) placed on the arms and the trunk from three volitional reaching motions representative of functional daily activities. We collected a dataset of eleven stroke participants performing a subset of FMA-UE, and three reaching motions. The FMA-UE of each participant was assessed by an occupational therapist providing the labeled score for the training data. The estimator was trained on windowed data during FMA-UE motions and was able to make score estimates from reaching motions. Through leave-one-subject-out cross validation, the estimator achieved a normalized RMSE of 7$\% $, which is comparable to or below the established minimal clinically important difference and minimal detectable change of FMA-UE of post-stroke individuals. Comparison experiments of various model designs also revealed the importance of trunk-based features inspired by compensation strategies common post stroke and features extracted from the hand sensor. The proposed estimator has the potential to broaden the possibility of automatic assessment via wearable sensors.

In settings where users both need high accuracy and are time-pressured, such as doctors working in emergency rooms, we want to provide AI assistance that both increases decision accuracy and reduces decision-making time. Current literature focusses on how users interact with AI assistance when there is no time pressure, finding that different AI assistances have different benefits: some can reduce time taken while increasing overreliance on AI, while others do the opposite. The precise benefit can depend on both the user and task. In time-pressured scenarios, adapting when we show AI assistance is especially important: relying on the AI assistance can save time, and can therefore be beneficial when the AI is likely to be right. We would ideally adapt what AI assistance we show depending on various properties (of the task and of the user) in order to best trade off accuracy and time. We introduce a study where users have to answer a series of logic puzzles. We find that time pressure affects how users use different AI assistances, making some assistances more beneficial than others when compared to no-time-pressure settings. We also find that a user’s overreliance rate is a key predictor of their behaviour: overreliers and not-overreliers use different AI assistance types differently. We find marginal correlations between a user’s overreliance rate (which is related to the user’s trust in AI recommendations) and their personality traits (Big Five Personality traits). Overall, our work suggests that AI assistances have different accuracy-time tradeoffs when people are under time pressure compared to no time pressure, and we explore how we might adapt AI assistances in this setting.

Discount regularization, using a shorter planning horizon when calculating the optimal policy, is a popular choice to avoid overfitting when faced with sparse or noisy data. It is commonly interpreted as de-emphasizing or ignoring delayed effects. In this paper, we prove two alternative views of discount regularization that expose unintended consequences and motivate novel regularization methods. In model-based RL, planning under a lower discount factor acts like a prior with stronger regularization on state-action pairs with more transition data. This leads to poor performance when the transition matrix is estimated from data sets with uneven amounts of data across state-action pairs. In model-free RL, discount regularization equates to planning using a weighted average Bellman update, where the agent plans as if the values of all state-action pairs are closer than implied by the data. Our equivalence theorems motivate simple methods that generalize discount regularization by setting parameters locally for individual state-action pairs rather than globally. We demonstrate the failures of discount regularization and how we remedy them using our state-action-specific methods across empirical examples with both tabular and continuous state spaces.

While dental disease is largely preventable, professional advice on optimal oral hygiene practices is often forgotten or abandoned by patients. Therefore patients may benefit from timely and personalized encouragement to engage in oral self-care behaviors. In this paper, we develop an online reinforcement learning (RL) algorithm for use in optimizing the delivery of mobile-based prompts to encourage oral hygiene behaviors. One of the main challenges in developing such an algorithm is ensuring that the algorithm considers the impact of current actions on the effectiveness of future actions (i.e., delayed effects), especially when the algorithm has been designed to run stably and autonomously in a constrained, real-world setting characterized by highly noisy, sparse data. We address this challenge by designing a quality reward that maximizes the desired health outcome (i.e., high-quality brushing) while minimizing user burden. We also highlight a procedure for optimizing the hyperparameters of the reward by building a simulation environment test bed and evaluating candidates using the test bed. The RL algorithm discussed in this paper will be deployed in Oralytics. To the best of our knowledge, Oralytics is the first mobile health study utilizing an RL algorithm designed to prevent dental disease by optimizing the delivery of motivational messages supporting oral self-care behaviors.

Discount regularization, using a shorter planning horizon when calculating the optimal policy, is a popular choice to restrict planning to a less complex set of policies when estimating an MDP from sparse or noisy data (Jiang et al., 2015). It is commonly understood that discount regularization functions by de-emphasizing or ignoring delayed effects. In this paper, we reveal an alternate view of discount regularization that exposes unintended consequences. We demonstrate that planning under a lower discount factor produces an identical optimal policy to planning using any prior on the transition matrix that has the same distribution for all states and actions. In fact, it functions like a prior with stronger regularization on state-action pairs with more transition data. This leads to poor performance when the transition matrix is estimated from data sets with uneven amounts of data across state-action pairs. Our equivalence theorem leads to an explicit formula to set regularization parameters locally for individual state-action pairs rather than globally. We demonstrate the failures of discount regularization and how we remedy them using our state-action-specific method across simple empirical examples as well as a medical cancer simulator.

Off-policy Evaluation (OPE) methods are crucial tools for evaluating policies in high-stakes domains such as healthcare, where direct deployment is often infeasible, unethical, or expensive. When deployment environments are expected to undergo changes (that is, dataset shifts), it is important for OPE methods to perform robust evaluation of the policies amidst such changes. Existing approaches consider robustness against a large class of shifts that can arbitrarily change any observable property of the environment. This often results in highly pessimistic estimates of the utilities, thereby invalidating policies that might have been useful in deployment. In this work, we address the aforementioned problem by investigating how domain knowledge can help provide more realistic estimates of the utilities of policies. We leverage human inputs on which aspects of the environments may plausibly change, and adapt the OPE methods to only consider shifts on these aspects. Specifically, we propose a novel framework, Robust OPE (ROPE), which considers shifts on a subset of covariates in the data based on user inputs, and estimates worst-case utility under these shifts. We then develop computationally efficient algorithms for OPE that are robust to the aforementioned shifts for contextual bandits and Markov decision processes. We also theoretically analyze the sample complexity of these algorithms. Extensive experimentation with synthetic and real world datasets from the healthcare domain demonstrates that our approach not only captures realistic dataset shifts accurately, but also results in less pessimistic policy evaluations.

Bayesian neural networks (BNNs) combine the expressive power of deep learning with the advantages of Bayesian formalism. In recent years, the analysis of wide, deep BNNs has provided theoretical insight into their priors and posteriors. However, we have no analogous insight into their posteriors under approximate inference. In this work, we show that mean-field variational inference entirely fails to model the data when the network width is large and the activation function is odd. Specifically, for fully-connected BNNs with odd activation functions and a homoscedastic Gaussian likelihood, we show that the optimal mean-field variational posterior predictive (i.e., function space) distribution converges to the prior predictive distribution as the width tends to infinity. We generalize aspects of this result to other likelihoods. Our theoretical results are suggestive of underfitting behavior previously observered in BNNs. While our convergence bounds are non-asymptotic and constants in our analysis can be computed, they are currently too loose to be applicable in standard training regimes. Finally, we show that the optimal approximate posterior need not tend to the prior if the activation function is not odd, showing that our statements cannot be generalized arbitrarily.

Recent years have seen a surge of interest in the field of explainable AI (XAI), with a plethora of algorithms proposed in the literature. However, a lack of consensus on how to evaluate XAI hinders the advancement of the field. We highlight that XAI is not a monolithic set of technologies---researchers and practitioners have begun to leverage XAI algorithms to build XAI systems that serve different usage contexts, such as model debugging and decision-support. Algorithmic research of XAI, however, often does not account for these diverse downstream usage contexts, resulting in limited effectiveness or even unintended consequences for actual users, as well as difficulties for practitioners to make technical choices. We argue that one way to close the gap is to develop evaluation methods that account for different user requirements in these usage contexts. Towards this goal, we introduce a perspective of contextualized XAI evaluation by considering the relative importance of XAI evaluation criteria for prototypical usage contexts of XAI. To explore the context dependency of XAI evaluation criteria, we conduct two survey studies, one with XAI topical experts and another with crowd workers. Our results urge for responsible AI research with usage-informed evaluation practices, and provide a nuanced understanding of user requirements for XAI in different usage contexts.

Many researchers and policymakers have expressed excitement about algorithmic explanations enabling more fair and responsible decision-making. However, recent experimental studies have found that explanations do not always improve human use of algorithmic advice. In this study, we shed light on how people interpret and respond to counterfactual explanations (CFEs)---explanations that show how a model's output would change with marginal changes to its input(s)---in the context of pretrial risk assessment instruments (PRAIs). We ran think-aloud trials with eight sitting U.S. state court judges, providing them with recommendations from a PRAI that includes CFEs. We found that the CFEs did not alter the judges' decisions. At first, judges misinterpreted the counterfactuals as real---rather than hypothetical---changes to defendants. Once judges understood what the counterfactuals meant, they ignored them, stating their role is only to make decisions regarding the actual defendant in question. The judges also expressed a mix of reasons for ignoring or following the advice of the PRAI without CFEs. These results add to the literature detailing the unexpected ways in which people respond to algorithms and explanations. They also highlight new challenges associated with improving human-algorithm collaborations through explanations.

Concept bottleneck models (CBMs) enhance the interpretability of their predictions by first predicting high-level concepts given features, and subsequently predicting outcomes on the basis of these concepts. Recently, it was demonstrated that training the label predictor directly on the probabilities produced by the concept predictor as opposed to the ground-truth concepts, improves label predictions. However, this results in corruptions in the concept predictions that impact the concept accuracy as well as our ability to intervene on the concepts -- a key proposed benefit of CBMs. In this work, we investigate and address two issues with CBMs that cause this disparity in performance: having an insufficient concept set and using inexpressive concept predictor. With our modifications, CBMs become competitive in terms of predictive performance, with models that otherwise leak additional information in the concept probabilities, while having dramatically increased concept accuracy and intervention accuracy.

Many reinforcement learning (RL) applications have combinatorial action spaces, where each action is a composition of sub-actions. A standard RL approach ignores this inherent factorization structure, resulting in a potential failure to make meaningful inferences about rarely observed sub-action combinations; this is particularly problematic for offline settings, where data may be limited. In this work, we propose a form of linear Q-function decomposition induced by factored action spaces. We study the theoretical properties of our approach, identifying scenarios where it is guaranteed to lead to zero bias when used to approximate the Q-function. Outside the regimes with theoretical guarantees, we show that our approach can still be useful because it leads to better sample efficiency without necessarily sacrificing policy optimality, allowing us to achieve a better bias-variance trade-off. Across several offline RL problems using simulators and real-world datasets motivated by healthcare, we demonstrate that incorporating factored action spaces into value-based RL can result in better-performing policies. Our approach can help an agent make more accurate inferences within underexplored regions of the state-action space when applying RL to observational datasets.

Bayesian Neural Networks with Latent Variables (BNN+LVs) capture predictive uncertainty by explicitly modeling model uncertainty (via priors on network weights) and environmental stochasticity (via a latent input noise variable). In this work, we first show that BNN+LV suffers from a serious form of non-identifiability: explanatory power can be transferred between the model parameters and latent variables while fitting the data equally well. We demonstrate that as a result, in the limit of infinite data, the posterior mode over the network weights and latent variables is asymptotically biased away from the ground-truth. Due to this asymptotic bias, traditional inference methods may in practice yield parameters that generalize poorly and misestimate uncertainty. Next, we develop a novel inference procedure that explicitly mitigates the effects of likelihood non-identifiability during training and yields high-quality predictions as well as uncertainty estimates. We demonstrate that our inference method improves upon benchmark methods across a range of synthetic and real data-sets.

Major depressive disorder is a debilitating disease affecting 264 million people worldwide. While many antidepressant medications are available, few clinical guidelines support choosing among them. Decision support tools (DSTs) embodying machine learning models may help improve the treatment selection process, but often fail in clinical practice due to poor system integration.

In representation learning, there has been recent interest in developing algorithms to disentangle the ground-truth generative factors behind a dataset, and metrics to quantify how fully this occurs. However, these algorithms and metrics often assume that both representations and ground-truth factors are flat, continuous, and factorized, whereas many real-world generative processes involve rich hierarchical structure, mixtures of discrete and continuous variables with dependence between them, and even varying intrinsic dimensionality. In this work, we develop benchmarks, algorithms, and metrics for learning such hierarchical representations.

Importance sampling-based estimators for off-policy evaluation (OPE) are valued for their simplicity, unbiasedness, and reliance on relatively few assumptions. However, the variance of these estimators is often high, especially when trajectories are of different lengths. In this work, we introduce Omitting-States-Irrelevant-to-Return Importance Sampling (OSIRIS), an estimator which reduces variance by strategically omitting likelihood ratios associated with certain states. We formalize the conditions under which OSIRIS is unbiased and has lower variance than ordinary importance sampling, and we demonstrate these properties empirically.

In the predict-then-optimize framework, the objective is to train a predictive model, mapping from environment features to parameters of an optimization problem, which maximizes decision quality when the optimization is subsequently solved. Recent work on decision-focused learning shows that embedding the optimization problem in the training pipeline can improve decision quality and help generalize better to unseen tasks compared to relying on an intermediate loss function for evaluating prediction quality. We study the predict-then-optimize framework in the context of sequential decision problems (formulated as MDPs) that are solved via reinforcement learning. In particular, we are given environment features and a set of trajectories from training MDPs, which we use to train a predictive model that generalizes to unseen test MDPs without trajectories. Two significant computational challenges arise in applying decision-focused learning to MDPs: (i) large state and action spaces make it infeasible for existing techniques to differentiate through MDP problems, and (ii) the high-dimensional policy space, as parameterized by a neural network, makes differentiating through a policy expensive. We resolve the first challenge by sampling provably unbiased derivatives to approximate and differentiate through optimality conditions, and the second challenge by using a low-rank approximation to the high-dimensional sample-based derivatives. We implement both Bellman-based and policy gradient-based decision-focused learning on three different MDP problems with missing parameters, and show that decision-focused learning performs better in generalization to unseen tasks.

Abstract This article describes a simple PCFG induction model with a fixed category domain that predicts a large majority of attested constituent boundaries, and predicts labels consistent with nearly half of attested constituent labels on a standard evaluation data set of child-directed speech. The article then explores the idea that the difference between simple grammars exhibited by child learners and fully recursive grammars exhibited by adult learners may be an effect of increasing working memory capacity, where the shallow grammars are constrained images of the recursive grammars. An implementation of these memory bounds as limits on center embedding in a depth-specific transform of a recursive grammar yields a significant improvement over an equivalent but unbounded baseline, suggesting that this arrangement may indeed confer a learning advantage.

Deep models have advanced prediction in many domains, but their lack of interpretability remains a key barrier to the adoption in many real world applications. There exists a large body of work aiming to help humans understand these black box functions to varying levels of granularity – for example, through distillation, gradients, or adversarial examples. These methods however, all tackle interpretability as a separate process after training. In this work, we take a different approach and explicitly regularize deep models so that they are well-approximated by processes that humans can step through in little time. Specifically, we train several families of deep neural networks to resemble compact, axis-aligned decision trees without significant compromises in accuracy. The resulting axis-aligned decision functions uniquely make tree regularized models easy for humans to interpret. Moreover, for situations in which a single, global tree is a poor estimator, we introduce a regional tree regularizer that encourages the deep model to resemble a compact, axis-aligned decision tree in predefined, human-interpretable contexts. Using intuitive toy examples, benchmark image datasets, and medical tasks for patients in critical care and with HIV, we demonstrate that this new family of tree regularizers yield models that are easier for humans to simulate than L1 or L2 penalties without sacrificing predictive power.

Ensembles depend on diversity for improved performance. Many ensemble training methods, therefore, attempt to optimize for diversity, which they almost always define in terms of differences in training set predictions. In this paper, however, we demonstrate the diversity of predictions on the training set does not necessarily imply diversity under mild covariate shift, which can harm generalization in practical settings. To address this issue, we introduce a new diversity metric and associated method of training ensembles of models that extrapolate differently on local patches of the data manifold. Across a variety of synthetic and real-world tasks, we find that our method improves generalization and diversity in qualitatively novel ways, especially under data limits and covariate shift.

The lack of interpretability remains a barrier to adopting deep neural networks across many safety-critical domains. Tree regularization was recently proposed to encourage a deep neural network's decisions to resemble those of a globally compact, axis-aligned decision tree. However, it is often unreasonable to expect a single tree to predict well across all possible inputs. In practice, doing so could lead to neither interpretable nor performant optima. To address this issue, we propose regional tree regularization – a method that encourages a deep model to be well-approximated by several separate decision trees specific to predefined regions of the input space. Across many datasets, including two healthcare applications, we show our approach delivers simpler explanations than other regularization schemes without compromising accuracy. Specifically, our regional regularizer finds many more “desirable” optima compared to global analogues.

Many medical decision-making tasks can be framed as partially observed Markov decision processes (POMDPs). However, prevailing two-stage approaches that first learn a POMDP and then solve it often fail because the model that best fits the data may not be well suited for planning. We introduce a new optimization objective that (a) produces both high-performing policies and high-quality generative models, even when some observations are irrelevant for planning, and (b) does so in batch off-policy settings that are typical in healthcare, when only retrospective data is available. We demonstrate our approach on synthetic examples and a challenging medical decision-making problem.

Supervised topic models are often sought to balance prediction quality and interpretability. However, when models are (inevitably) misspecified, standard approaches rarely deliver on both. We introduce a novel approach, the prediction-focused topic model, that uses the supervisory signal to retain only vocabulary terms that improve, or at least do not hinder, prediction performance. By removing terms with irrelevant signal, the topic model is able to learn task-relevant, coherent topics. We demonstrate on several data sets that compared to existing approaches, prediction-focused topic models learn much more coherent topics while maintaining competitive predictions.

Off-policy evaluation in reinforcement learning offers the chance of using observational data to improve future outcomes in domains such as healthcare and education, but safe deployment in high stakes settings requires ways of assessing its validity. Traditional measures such as confidence intervals may be insufficient due to noise, limited data and confounding. In this paper we develop a method that could serve as a hybrid human-AI system, to enable human experts to analyze the validity of policy evaluation estimates. This is accomplished by highlighting observations in the data whose removal will have a large effect on the OPE estimate, and formulating a set of rules for choosing which ones to present to domain experts for validation. We develop methods to compute exactly the influence functions for fitted Q-evaluation with two different function classes: kernel-based and linear least squares, as well as importance sampling methods. Experiments on medical simulations and real-world intensive care unit data demonstrate that our method can be used to identify limitations in the evaluation process and make evaluation more robust.

We present two elegant solutions for modeling continuous-time dynamics, in a novel model-based reinforcement learning (RL) framework for semi-Markov decision processes (SMDPs), using neural ordinary differential equations (ODEs). Our models accurately characterize continuous-time dynamics and enable us to develop high-performing policies using a small amount of data. We also develop a model-based approach for optimizing time schedules to reduce interaction rates with the environment while maintaining the near-optimal performance, which is not possible for model-free methods. We experimentally demonstrate the efficacy of our methods across various continuous-time domains.

Domains where supervised models are deployed often come with task-specific constraints, such as prior expert knowledge on the ground-truth function, or desiderata like safety and fairness. We introduce a novel probabilistic framework for reasoning with such constraints and formulate a prior that enables us to effectively incorporate them into Bayesian neural networks (BNNs), including a variant that can be amortized over tasks. The resulting Output-Constrained BNN (OC-BNN) is fully consistent with the Bayesian framework for uncertainty quantification and is amenable to black-box inference. Unlike typical BNN inference in uninterpretable parameter space, OC-BNNs widen the range of functional knowledge that can be incorporated, especially for model users without expertise in machine learning. We demonstrate the efficacy of OC-BNNs on real-world datasets, spanning multiple domains such as healthcare, criminal justice, and credit scoring.

Bayesian neural networks (BNNs) are flexible function priors well-suited to situations in which data are scarce and uncertainty must be quantified. Yet, common weight priors are able to encode little functional knowledge and can behave in undesirable ways. We present a novel prior over radial basis function networks (RBFNs) that allows for independent specification of functional amplitude variance and lengthscale (i.e., smoothness), where the inverse lengthscale corresponds to the concentration of radial basis functions. When the lengthscale is uniform over the input space, we prove consistency and approximate variance stationarity. This is in contrast to common BNN priors, which are highly nonstationary. When the input dependence of the lengthscale is unknown, we show how it can be inferred. We compare this model’s behavior to standard BNNs and Gaussian processes using synthetic and real examples.

Unsupervised PCFG inducers hypothesize sets of compact context-free rules as explanations for sentences.These models not only provide tools for low-resource languages, but also play an important role in modeling language acquisition (Bannard et al., 2009;Abend et al., 2017).However, current PCFG induction models, using word tokens as input, are unable to incorporate semantics and morphology into induction, and may encounter issues of sparse vocabulary when facing morphologically rich languages.This paper describes a neural PCFG inducer which employs context embeddings (Peters et al., 2018) in a normalizing flow model (Dinh et al., 2015) to extend PCFG induction to use semantic and morphological information 1 .Linguistically motivated similarity penalty and categorical distance constraints are imposed on the inducer as regularization.Experiments show that the PCFG induction model with normalizing flow produces grammars with state-of-the-art accuracy on a variety of different languages.Ablation further shows a positive effect of normalizing flow, context embeddings and proposed regularizers.

Recent years have seen a boom in interest in interpretable machine learning systems built on models that can be understood, at least to some degree, by domain experts. However, exactly what kinds of models are truly human-interpretable remains poorly understood. This work advances our understanding of precisely which factors make models interpretable in the context of decision sets, a specific class of logic-based model. We conduct carefully controlled human-subject experiments in two domains across three tasks based on human-simulatability through which we identify specific types of complexity that affect performance more heavily than others-trends that are consistent across tasks and domains. These results can inform the choice of regularizers during optimization to learn more interpretable models, and their consistency suggests that there may exist common design principles for interpretable machine learning systems.

We consider a model-based approach to perform batch off-policy evaluation in reinforcement learning. Our method takes a mixture-of-experts approach to combine parametric and non-parametric models of the environment such that the final value estimate has the least expected error. We do so by first estimating the local accuracy of each model and then using a planner to select which model to use at every time step as to minimize the return error estimate along entire trajectories. Across a variety of domains, our mixture-based approach outperforms the individual models alone as well as state-of-the-art importance sampling-based estimators.

AI agents support high stakes decision-making processes from driving cars to prescribing drugs, making it increasingly important for human users to understand their behavior. Policy summarization methods aim to convey strengths and weaknesses of such agents by demonstrating their behavior in a subset of informative states. Some policy summarization methods extract a summary that optimizes the ability to reconstruct the agent's policy under the assumption that users will deploy inverse reinforcement learning. In this paper, we explore the use of different models for extracting summaries. We introduce an imitation learning-based approach to policy summarization; we demonstrate through computational simulations that a mismatch between the model used to extract a summary and the model used to reconstruct the policy results in worse reconstruction quality; and we demonstrate through a human-subject study that people use different models to reconstruct policies in different contexts, and that matching the summary extraction model to these can improve performance. Together, our results suggest that it is important to carefully consider user models in policy summarization.

We introduce a novel apprenticeship learning algorithm to learn an expert's underlying reward structure in off-policy model-free batch settings. Unlike existing methods that require hand-crafted features, on-policy evaluation, further data acquisition for evaluation policies or the knowledge of model dynamics, our algorithm requires only batch data (demonstrations) of the observed expert behavior. Such settings are common in many real-world tasks---health care, finance, or industrial process control---where accurate simulators do not exist and additional data acquisition is costly. We develop a transition-regularized imitation learning model to learn a rich feature representation and a near-expert initial policy that makes the subsequent batch inverse reinforcement learning process viable. We also introduce deep successor feature networks that perform off-policy evaluation to estimate feature expectations of candidate policies. Under the batch setting, our method achieves superior results on control benchmarks as well as a real clinical task of sepsis management in the Intensive Care Unit.

Standard reinforcement learning methods aim to master one way of solving a task whereas there may exist multiple near-optimal policies. Being able to identify this collection of near-optimal policies can allow a domain expert to efficiently explore the space of reasonable solutions. Unfortunately, existing approaches that quantify uncertainty over policies are not ultimately relevant to finding policies with qualitatively distinct behaviors. In this work, we formalize the difference between policies as a difference between the distribution of trajectories induced by each policy, which encourages diversity with respect to both state visitation and action choices. We derive a gradient-based optimization technique that can be combined with existing policy gradient methods to now identify diverse collections of well-performing policies. We demonstrate our approach on benchmarks and a healthcare task.

The promise of augmenting accurate predictions provided by modern neural networks with well-calibrated predictive uncertainties has reinvigorated interest in Bayesian neural networks. However, model selection---even choosing the number of nodes---remains an open question. Poor choices can severely affect the quality of the produced uncertainties. In this paper, we explore continuous shrinkage priors, the horseshoe, and the regularized horseshoe distributions, for model selection in Bayesian neural networks. When placed over node pre-activations and coupled with appropriate variational approximations, we find that the strong shrinkage provided by the horseshoe is effective at turning off nodes that do not help explain the data. We demonstrate that our approach finds compact network structures even when the number of nodes required is grossly over-estimated. Moreover, the model selection over the number of nodes does not come at the expense of predictive or computational performance; in fact, we learn smaller networks with comparable predictive performance to current approaches. These effects are particularly apparent in sample-limited settings, such as small data sets and reinforcement learning.

Bayesian Non-negative Matrix Factorization (BNMF) is a promising approach for understanding uncertainty and structure in matrix data. However, a large volume of applied work optimizes traditional non-Bayesian NMF objectives that fail to provide a principled understanding of the non-identifiability inherent in NMF---an issue ideally addressed by a Bayesian approach. Despite their suitability, current BNMF approaches have failed to gain popularity in an applied setting; they sacrifice flexibility in modeling for tractable computation, tend to get stuck in local modes, and can require many thousands of samples for meaningful uncertainty estimates. We address these issues through a particle-based variational approach to BNMF that only requires the joint likelihood to be differentiable for computational tractability, uses a novel transfer-based initialization technique to identify multiple modes in the posterior, and thus allows domain experts to inspect a small set of factorizations that faithfully represent the posterior. On several real datasets, we obtain better particle approximations to the BNMF posterior in less time than baselines and demonstrate the significant role that multimodality plays in NMF-related tasks.

Deep neural networks have proven remarkably effective at solving many classification problems, but have been criticized recently for two major weaknesses: the reasons behind their predictions are uninterpretable, and the predictions themselves can often be fooled by small adversarial perturbations. These problems pose major obstacles for the adoption of neural networks in domains that require security or transparency. In this work, we evaluate the effectiveness of defenses that differentiably penalize the degree to which small changes in inputs can alter model predictions. Across multiple attacks, architectures, defenses, and datasets, we find that neural networks trained with this input gradient regularization exhibit robustness to transferred adversarial examples generated to fool all of the other models. We also find that adversarial examples generated to fool gradient-regularized models fool all other models equally well, and actually lead to more "legitimate," interpretable misclassifications as rated by people (which we confirm in a human subject experiment). Finally, we demonstrate that regularizing input gradients makes them more naturally interpretable as rationales for model predictions. We conclude by discussing this relationship between interpretability and robustness in deep neural networks.

The lack of interpretability remains a key barrier to the adoption of deep models in many applications. In this work, we explicitly regularize deep models so human users might step through the process behind their predictions in little time. Specifically, we train deep time-series models so their class-probability predictions have high accuracy while being closely modeled by decision trees with few nodes. Using intuitive toy examples as well as medical tasks for treating sepsis and HIV, we demonstrate that this new tree regularization yields models that are easier for humans to simulate than simpler L1 or L2 penalties without sacrificing predictive power.

Tensor decomposition methods are popular tools for learning latent variables given only lowerorder moments of the data. However, the standard assumption is that we have sufficient data to estimate these moments to high accuracy. In this work, we consider the case in which certain dimensions of the data are not always observed–common in applied settings, where not all measurements may be taken for all observations–resulting in moment estimates of varying quality. We derive a weighted tensor decomposition approach that is computationally as efficient as the non-weighted approach, and demonstrate that it outperforms methods that do not appropriately leverage these less-observed dimensions.

Supervisory signals can help topic models discover low-dimensional data representations which are useful for a specific prediction task. We propose a framework for training supervised latent Dirichlet allocation that balances two goals: faithful generative explanations of high-dimensional data and accurate prediction of associated class labels. Existing approaches fail to balance these goals by not properly handling a fundamental asymmetry: the intended application is always predicting labels from data, not data from labels. Our new prediction-constrained objective for training generative models coherently integrates supervisory signals even when only a small fraction of training examples are labeled. We demonstrate improved prediction quality compared to previous supervised topic models, achieving results competitive with high-dimensional logistic regression on text analysis and electronic health records tasks while simultaneously learning interpretable topics.

There have been several recent attempts to improve the accuracy of grammar induction systems by bounding the recursive complexity of the induction model (Ponvert et al., 2011;Noji and Johnson, 2016;Shain et al., 2016;Jin et al., 2018).Modern depth-bounded grammar inducers have been shown to be more accurate than early unbounded PCFG inducers, but this technique has never been compared against unbounded induction within the same system, in part because most previous depthbounding models are built around sequence models, the complexity of which grows exponentially with the maximum allowed depth.The present work instead applies depth bounds within a chart-based Bayesian PCFG inducer (Johnson et al., 2007b), where bounding can be switched on and off, and then samples trees with and without bounding.1 Results show that depth-bounding is indeed significantly effective in limiting the search space of the inducer and thereby increasing the accuracy of the resulting parsing model.Moreover, parsing results on English, Chinese and German show that this bounded model with a new inference technique is able to produce parse trees more accurately than or competitively with state-ofthe-art constituency-based grammar induction models.

Bayesian neural networks with latent variables are scalable and flexible probabilistic models: they account for uncertainty in the estimation of the network weights and, by making use of latent variables, can capture complex noise patterns in the data. Using these models we show how to perform and utilize a decomposition of uncertainty in aleatoric and epistemic components for decision making purposes. This allows us to successfully identify informative points for active learning of functions with heteroscedastic and bimodal noise. Using the decomposition we further define a novel risk-sensitive criterion for reinforcement learningto identify policies that balance expected cost, model-bias and noise aversion.

Bayesian Neural Networks (BNNs) have recently received increasing attention for their ability to provide well-calibrated posterior uncertainties. However, model selection—even choosing the number of nodes—remains an open question. Recent work has proposed the use of a horseshoe prior over node pre-activations of a Bayesian neural network, which effectively turns off nodes that do not help explain the data. In this work, we propose several modeling and inference advances that consistently improve the compactness of the model learned while maintaining predictive performance, especially in smaller-sample settings including reinforcement learning.

We often desire our models to be interpretable as well as accurate. Prior work on optimizing models for interpretability has relied on easy-to-quantify proxies for interpretability, such as sparsity or the number of operations required. In this work, we optimize for interpretability by directly including humans in the optimization loop. We develop an algorithm that minimizes the number of user studies to find models that are both predictive and interpretable and demonstrate our approach on several data sets. Our human subjects results show trends towards different proxy notions of interpretability on different datasets, which suggests that different proxies are preferred on different tasks.

We study the problem of off-policy policy evaluation (OPPE) in RL. In contrast to prior work, we consider how to estimate both the individual policy value and average policy value accurately. We draw inspiration from recent work in causal reasoning, and propose a new finite sample generalization error bound for value estimates from MDP models. Using this upper bound as an objective, we develop a learning algorithm of an MDP model with a balanced representation, and show that our approach can yield substantially lower MSE in common synthetic benchmarks and a HIV treatment simulation domain.

There has been recent interest in applying cognitively- or empirically-motivated bounds on recursion depth to limit the search space of grammar induction models (Ponvert et al., 2011; Noji and Johnson, 2016; Shain et al., 2016). This work extends this depth-bounding approach to probabilistic context-free grammar induction (DB-PCFG), which has a smaller parameter space than hierarchical sequence models, and therefore more fully exploits the space reductions of depth-bounding. Results for this model on grammar acquisition from transcribed child-directed speech and newswire text exceed or are competitive with those of other models when evaluated on parse accuracy. Moreover, grammars acquired from this model demonstrate a consistent use of category labels, something which has not been demonstrated by other acquisition models.

PhenoLines is a visual analysis tool for the interpretation of disease subtypes, derived from the application of topic models to clinical data. Topic models enable one to mine cross-sectional patient comorbidity data (e.g., electronic health records) and construct disease subtypes-each with its own temporally evolving prevalence and co-occurrence of phenotypes-without requiring aligned longitudinal phenotype data for all patients. However, the dimensionality of topic models makes interpretation challenging, and de facto analyses provide little intuition regarding phenotype relevance or phenotype interrelationships. PhenoLines enables one to compare phenotype prevalence within and across disease subtype topics, thus supporting subtype characterization, a task that involves identifying a proposed subtype's dominant phenotypes, ages of effect, and clinical validity. We contribute a data transformation workflow that employs the Human Phenotype Ontology to hierarchically organize phenotypes and aggregate the evolving probabilities produced by topic models. We introduce a novel measure of phenotype relevance that can be used to simplify the resulting topology. The design of PhenoLines was motivated by formative interviews with machine learning and clinical experts. We describe the collaborative design process, distill high-level tasks, and report on initial evaluations with machine learning experts and a medical domain expert. These results suggest that PhenoLines demonstrates promising approaches to support the characterization and optimization of topic models.

An intriguing application of transfer learning emerges when tasks arise with similar, but not identical, dynamics. Hidden Parameter Markov Decision Processes (HiP-MDP) embed these tasks into a low-dimensional space; given the embedding parameters one can identify the MDP for a particular task. However, the original formulation of HiP-MDP had a critical flaw: the embedding uncertainty was modeled independently of the agent's state uncertainty, requiring an arduous training procedure. In this work, we apply a Gaussian Process latent variable model to jointly model the dynamics and the embedding, leading to a more elegant formulation, one that allows for better uncertainty quantification and thus more robust transfer.

Expressive classifiers such as neural networks are among the most accurate supervised learning methods in use today, but their opaque decision boundaries make them difficult to trust in critical applications. We propose a method to explain the predictions of any differentiable model via the gradient of the class label with respect to the input (which provides a normal to the decision boundary). Not only is this approach orders of magnitude faster at identifying input dimensions of high sensitivity than sample-based perturbation methods (e.g. LIME), but it also lends itself to efficiently discovering multiple qualitatively different decision boundaries as well as decision boundaries that are consistent with expert annotation. On multiple datasets, we show our approach generalizes much better when test conditions differ from those in training.

We introduce a new formulation of the Hidden Parameter Markov Decision Process (HiP-MDP), a framework for modeling families of related tasks using low-dimensional latent embeddings. Our new framework correctly models the joint uncertainty in the latent parameters and the state space. We also replace the original Gaussian Process-based model with a Bayesian Neural Network, enabling more scalable inference. Thus, we expand the scope of the HiP-MDP to applications with higher dimensions and more complex dynamics.

BACKGROUND: The widespread adoption of electronic health records allows us to ask evidence-based questions about the need for and benefits of specific clinical interventions in critical-care settings across large populations. OBJECTIVE: We investigated the prediction of vasopressor administration and weaning in the intensive care unit. Vasopressors are commonly used to control hypotension, and changes in timing and dosage can have a large impact on patient outcomes. MATERIALS AND METHODS: We considered a cohort of 15 695 intensive care unit patients without orders for reduced care who were alive 30 days post-discharge. A switching-state autoregressive model (SSAM) was trained to predict the multidimensional physiological time series of patients before, during, and after vasopressor administration. The latent states from the SSAM were used as predictors of vasopressor administration and weaning. RESULTS: The unsupervised SSAM features were able to predict patient vasopressor administration and successful patient weaning. Features derived from the SSAM achieved areas under the receiver operating curve of 0.92, 0.88, and 0.71 for predicting ungapped vasopressor administration, gapped vasopressor administration, and vasopressor weaning, respectively. We also demonstrated many cases where our model predicted weaning well in advance of a successful wean. CONCLUSION: Models that used SSAM features increased performance on both predictive tasks. These improvements may reflect an underlying, and ultimately predictive, latent state detectable from the physiological time series.

We present a machine learning algorithm for building classifiers that are comprised of a small number of short rules. These are restricted disjunctive normal form models. An example of a classifier of this form is as follows: If $X$ satisfies (condition $A$ AND condition $B$) OR (condition $C$) OR $\cdots$, then $Y=1$. Models of this form have the advantage of being interpretable to human experts since they produce a set of rules that concisely describe a specific class. We present two probabilistic models with prior parameters that the user can set to encourage the model to have a desired size and shape, to conform with a domain-specific definition of interpretability. We provide a scalable MAP inference approach and develop theoretical bounds to reduce computation by iteratively pruning the search space. We apply our method (Bayesian Rule Sets -- BRS) to characterize and predict user behavior with respect to in-vehicle context-aware personalized recommender systems. Our method has a major advantage over classical associative classification methods and decision trees in that it does not greedily grow the model.

Method of moment estimators exhibit appealing statistical properties, such as asymptotic unbiasedness, for nonconvex problems. However, they typically require a large number of samples and are extremely sensitive to model misspecification. In this paper, we apply the framework of M-estimation to develop both a generalized method of moments procedure and a principled method for regularization. Our proposed M-estimator obtains optimal sample efficiency rates (in the class of moment-based estimators) and the same well-known rates on prediction accuracy as other spectral estimators. It also makes it straightforward to incorporate regularization into the sample moment conditions. We demonstrate empirically the gains in sample efficiency from our approach on hidden Markov models.

This paper presents a new memory-bounded left-corner parsing model for unsupervised raw-text syntax induction, using unsupervised hierarchical hidden Markov models (UHHMM). We deploy this algorithm to shed light on the extent to which human language learners can discover hierarchical syntax through distributional statistics alone, by modeling two widely-accepted features of human language acquisition and sentence processing that have not been simultaneously modeled by any existing grammar induction algorithm: (1) a left-corner parsing strategy and (2) limited working memory capacity. To model realistic input to human language learners, we evaluate our system on a corpus of child-directed speech rather than typical newswire corpora. Results beat or closely match those of three competing systems.

A Rule Set model consists of a small number of short rules for interpretable classification, where an instance is classified as positive if it satisfies at least one of the rules. The rule set provides reasons for predictions, and also descriptions of a particular class. We present a Bayesian framework for learning Rule Set models, with prior parameters that the user can set to encourage the model to have a desired size and shape in order to conform with a domain-specific definition of interpretability. We use an efficient inference approach for searching for the MAP solution and provide theoretical bounds to reduce computation. We apply Rule Set models to ten UCI data sets and compare the performance with other interpretable and non-interpretable models.

Astronomers and telescope operators must make decisions about what to observe given limited telescope time. To optimize this decision-making process, we present a batch, cost-sensitive, active learning approach that exploits structure in the unlabeled dataset, accounts for label uncertainty, and minimizes annotation costs. We first cluster the unlabeled instances in feature space. We next introduce an uncertainty-reducing selection criterion that encourages the batch of selected instances to span multiple clusters, in addition to taking into account annotation cost. Finally, we extend this criterion to incorporate the fact that nearby astronomical objects may be observed at the same time. On two large astronomical data sets, our approach balances the trade-offs among FOV, aperture, and time cost and, therefore, helps astronomers design effective experiments.

Patients with developmental disorders, such as autism spectrum disorder (ASD), present with symptoms that change with time even if the named diagnosis remains fixed. For example, language impairments may present as delayed speech in a toddler and difficulty reading in a school-age child. Characterizing these trajectories is important for early treatment. However, deriving these trajectories from observational sources is challenging: electronic health records only reflect observations of patients at irregular intervals and only record what factors are clinically relevant at the time of observation. Meanwhile, caretakers discuss daily developments and concerns on social media. In this work, we present a fully unsupervised approach for learning disease trajectories from incomplete medical records and social media posts, including cases in which we have only a single observation of each patient. In particular, we use a dynamic topic model approach which embeds each disease trajectory as a path in $\mathbb{R}^D$. A Polya- gamma augmentation scheme is used to efficiently perform inference as well as incorporate multiple data sources. We learn disease trajectories from the electronic health records of 13,435 patients with ASD and the forum posts of 13,743 caretakers of children with ASD, deriving interesting clinical insights as well as good predictions.

Topic modeling is a powerful tool for uncovering latent structure in many domains, including medicine, finance, and vision. The goals for the model vary depending on the application: sometimes the discovered topics are used for prediction or another downstream task. In other cases, the content of the topic may be of intrinsic scientific interest. Unfortunately, even when one uses modern sparse techniques, discovered topics are often difficult to interpret due to the high dimensionality of the underlying space. To improve topic interpretability, we introduce Graph-Sparse LDA, a hierarchical topic model that uses knowledge of relationships between words (e.g., as encoded by an ontology). In our model, topics are summarized by a few latent concept-words from the underlying graph that explain the observed words. Graph-Sparse LDA recovers sparse, interpretable summaries on two real-world biomedical datasets while matching state-of-the-art prediction performance.

We present the Mind the Gap Model (MGM), an approach for interpretable feature extraction and selection. By placing interpretability criteria directly into the model, we allow for the model to both optimize parameters related to interpretability and to directly report a global set of distinguishable dimensions to assist with further data exploration and hypothesis generation. MGM extracts distinguishing features on real-world datasets of animal features, recipes ingredients, and disease co-occurrence. It also maintains or improves performance when compared to related approaches. We perform a user study with domain experts to show the MGM's ability to help with dataset exploration.

Making intelligent decisions from incomplete information is critical in many applications: for example, robots must choose actions based on imperfect sensors, and speech-based interfaces must infer a user's needs from noisy microphone inputs. What makes these tasks hard is that often we do not have a natural representation with which to model the domain and use for choosing actions; we must learn about the domain's properties while simultaneously performing the task. Learning a representation also involves trade-offs between modeling the data that we have seen previously and being able to make predictions about new data. This article explores learning representations of stochastic systems using Bayesian nonparametric statistics. Bayesian nonparametric methods allow the sophistication of a representation to scale gracefully with the complexity in the data. Our main contribution is a careful empirical evaluation of how representations learned using Bayesian nonparametric methods compare to other standard learning approaches, especially in support of planning and control. We show that the Bayesian aspects of the methods result in achieving state-of-the-art performance in decision making with relatively few samples, while the nonparametric aspects often result in fewer computations. These results hold across a variety of different techniques for choosing actions given a representation.

Accurate knowledge of a patient's disease state and trajectory is critical in a clinical setting. Modern electronic healthcare records contain an increasingly large amount of data, and the ability to automatically identify the factors that influence patient outcomes stand to greatly improve the efficiency and quality of care. We examined the use of latent variable models (viz. Latent Dirichlet Allocation) to decompose free-text hospital notes into meaningful features, and the predictive power of these features for patient mortality. We considered three prediction regimes: (1) baseline prediction, (2) dynamic (time-varying) outcome prediction, and (3) retrospective outcome prediction. In each, our prediction task differs from the familiar time-varying situation whereby data accumulates; since fewer patients have long ICU stays, as we move forward in time fewer patients are available and the prediction task becomes increasingly difficult. We found that latent topic-derived features were effective in determining patient mortality under three timelines: inhospital, 30 day post-discharge, and 1 year post-discharge mortality. Our results demonstrated that the latent topic features important in predicting hospital mortality are very different from those that are important in post-discharge mortality. In general, latent topic features were more predictive than structured features, and a combination of the two performed best. The time-varying models that combined latent topic features and baseline features had AUCs that reached 0.85, 0.80, and 0.77 for in-hospital, 30 day post-discharge and 1 year post-discharge mortality respectively. Our results agreed with other work suggesting that the first 24 hours of patient information are often the most predictive of hospital mortality. Retrospective models that used a combination of latent topic features and structured features achieved AUCs of 0.96, 0.82, and 0.81 for in-hospital, 30 day, and 1-year mortality prediction. Our work focuses on the dynamic (time-varying) setting because models from this regime could facilitate an on-going severity stratification system that helps direct care-staff resources and inform treatment strategies.

The batteries of many consumer products are both a substantial portion of the product's cost and commonly a first point of failure. Accurately predicting remaining battery life can lower costs by reducing unnecessary battery replacements. Unfortunately, battery dynamics are extremely complex, and we often lack the domain knowledge required to construct a model by hand. In this work, we take a data-driven approach and aim to learn a model of battery time-to-death from training data. Using a Dirichlet process prior over mixture weights, we learn an infinite mixture model for battery health. The Bayesian aspect of our model helps to avoid over-fitting while the nonparametric nature of the model allows the data to control the size of the model, preventing under-fitting. We demonstrate our model's effectiveness by making time-to-death predictions using real data from nickel-metal hydride battery packs.

Constructing models of mobile agents can be difficult without domain-specific knowledge. Parametric models flexible enough to capture all mobility patterns that an expert believes are possible are often large, requiring a great deal of training data. In contrast, nonparametric models are extremely flexible and can generalize well with relatively little training data. We propose modeling the mobility patterns of moving agents as a mixture of Gaussian processes (GP) with a Dirichlet process (DP) prior over mixture weights. The GP provides a flexible representation for each individual mobility pattern, while the DP assigns observed trajectories to particular mobility patterns. Both the GPs and the DP adjust the model's complexity based on available data, implicitly avoiding issues of over-fitting or under-fitting. We apply our model to a helicopter-based tracking task, where the mobility patterns of the tracked agents — cars — are learned from real data collected from taxis in the greater Boston area.

We consider reinforcement learning in partially observable domains where the agent can query an expert for demonstrations. Our nonparametric Bayesian approach combines model knowledge, inferred from expert information and independent exploration, with policy knowledge inferred from expert trajectories. We introduce priors that bias the agent towards models with both simple representations and simple policies, resulting in improved policy and model learning.

We often seek to identify co-occurring hidden features in a set of observations. The Indian Buffet Process (IBP) provides a non-parametric prior on the features present in each observation, but current inference techniques for the IBP often scale poorly. The collapsed Gibbs sampler for the IBP has a running time cubic in the number of observations, and the uncollapsed Gibbs sampler, while linear, is often slow to mix. We present a new linear-time collapsed Gibbs sampler for conjugate likelihood models and demonstrate its efficacy on large real-world datasets.

The Partially Observable Markov Decision Process (POMDP) framework has proven useful in planning domains that require balancing actions that increase an agents knowledge and actions that increase an agents reward. Unfortunately, most POMDPs are complex structures with a large number of parameters. In many realworld problems, both the structure and the parameters are difficult to specify from domain knowledge alone. Recent work in Bayesian reinforcement learning has made headway in learning POMDP models; however, this work has largely focused on learning the parameters of the POMDP model. We define an infinite POMDP (iPOMDP) model that does not require knowledge of the size of the state space; instead, it assumes that the number of visited states will grow as the agent explores its world and explicitly models only visited states. We demonstrate the iPOMDP utility on several standard problems.

Nonparametric Bayesian models provide a framework for flexible probabilistic modelling of complex datasets. Unfortunately, Bayesian inference methods often require high-dimensional averages and can be slow to compute, especially with the potentially unbounded representations associated with nonparametric models. We address the challenge of scaling nonparametric Bayesian inference to the increasingly large datasets found in real-world applications, focusing on the case of parallelising inference in the Indian Buffet Process (IBP). Our approach divides a large data set between multiple processors. The processors use message passing to compute likelihoods in an asynchronous, distributed fashion and to propagate statistics about the global Bayesian posterior. This novel MCMC sampler is the first parallel inference scheme for IBP-based models, scaling to datasets orders of magnitude larger than had previously been possible.

Spoken language is one of the most intuitive forms of interaction between humans and agents. Unfortunately, agents that interact with people using natural language often experience communication errors and do not correctly understand the user's intentions. Recent systems have successfully used probabilistic models of speech, language and user behaviour to generate robust dialogue performance in the presence of noisy speech recognition and ambiguous language choices, but decisions made using these probabilistic models are still prone to errors owing to the complexity of acquiring and maintaining a complete model of human language and behaviour. In this paper, a decision-theoretic model for human–robot interaction using natural language is described. The algorithm is based on the Partially Observable Markov Decision Process (POMDP), which allows agents to choose actions that are robust not only to uncertainty from noisy or ambiguous speech recognition but also unknown user models. Like most dialogue systems, a POMDP is defined by a large number of parameters that may be difficult to specify a priori from domain knowledge, and learning these parameters from the user may require an unacceptably long training period. An extension to the POMDP model is described that allows the agent to acquire a linguistic model of the user online, including new vocabulary and word choice preferences. The approach not only avoids a training period of constant questioning as the agent learns, but also allows the agent actively to query for additional information when its uncertainty suggests a high risk of mistakes. The approach is demonstrated both in simulation and on a natural language interaction system for a robotic wheelchair application.

Intelligent planning algorithms such as the Partially Observable Markov Decision Process (POMDP) have succeeded in dialog management applications [10, 11, 12] because they are robust to the inherent uncertainty of human interaction. Like all dialog planning systems, however, POMDPs require an accurate model of the user (e.g., what the user might say or want). POMDPs are generally specified using a large probabilistic model with many parameters. These parameters are difficult to specify from domain knowledge, and gathering enough data to estimate the parameters accurately a priori is expensive.In this paper, we take a Bayesian approach to learning the user model simultaneously with dialog manager policy. At the heart of our approach is an efficient incremental update algorithm that allows the dialog manager to replan just long enough to improve the current dialog policy given data from recent interactions. The update process has a relatively small computational cost, preventing long delays in the interaction. We are able to demonstrate a robust dialog manager that learns from interaction data, out-performing a hand-coded model in simulation and in a robotic wheelchair application.

We propose a fast approach for detecting collision- free swing-foot trajectories for legged locomotion over extreme terrains. Instead of simulating the swing trajectories and checking for collisions along them, our approach uses machine learning techniques to predict whether a swing trajectory is collision-free. Using a set of local terrain features, we apply supervised learning to train a classifier to predict collisions. Both in simulation and on a real quadruped platform, our results show that our classifiers can improve the accuracy of collision detection compared to a real-time geometric approach without significantly increasing the computation time.