We introduce TextWorld, a sandbox learning environment for the training and evaluation of RL agents on text-based games. TextWorld is a Python library that handles interactive play-through of text games, as well as backend functions like state tracking and reward assignment. It comes with a curated list of games whose features and challenges we have analyzed. More significantly, it enables users to handcraft or automatically generate new games. Its generative mechanisms give precise control over the difficulty, scope, and language of constructed games, and can be used to relax challenges inherent to commercial text games like partial observability and sparse rewards. By generating sets of varied but similar games, TextWorld can also be used to study generalization and transfer learning. We cast text-based games in the Reinforcement Learning formalism, use our framework to develop a set of benchmark games, and evaluate several baseline agents on this set and the curated list.

We present a Bayesian view of counterfactual risk minimization (CRM), also known as offline policy optimization from logged bandit feedback. Using PAC-Bayesian analysis, we derive a new generalization bound for the truncated IPS estimator. We apply the bound to a class of Bayesian policies, which motivates a novel, potentially data-dependent, regularization technique for CRM.

Bayesian methods in machine learning, such as Gaussian processes, have great advantages compared to other techniques. In particular, they provide estimates of the uncertainty associated with a prediction. Extending the Bayesian approach to deep architectures has remained a major challenge. Recent results connected deep feedforward neural networks with Gaussian processes, allowing training without backpropagation. This connection enables us to leverage a quantum algorithm designed for Gaussian processes and develop new algorithms for Bayesian deep learning on quantum computers. The properties of the kernel matrix in the Gaussian process ensure the efficient execution of the core component of the protocol, quantum matrix inversion, providing a polynomial speedup with respect to classical algorithm. Furthermore, we demonstrate the execution of the algorithm on contemporary quantum computers and analyze its robustness to realistic noise models.

Generative models are known to be difficult to assess. Recent works, especially on generative adversarial networks (GANs), produce good visual samples of varied categories of images. However, the validation of their quality is still difficult to define and there is no existing agreement on the best evaluation process. This paper aims at making a step toward an objective evaluation process for generative models. It presents a new method to assess a trained generative model by evaluating its capacity to teach a classification task to a discriminative model. Our approach evaluates generators on a testing set by using, as a proxy, a neural network classifier trained on generated samples. Neural networks classifier are known to be difficult to train on an unbalanced or biased dataset. We use this weakness as a proxy to evaluate generated data and hence generative models. Our assumption is that to train a successful neural network classifier, the training data should contain meaningful and varied information, that fit and capture the whole distribution of the testing dataset. By comparing results with different generated datasets we can classify generative models. The motivation of this approach is also to evaluate if generative models can help discriminative neural networks to learn, i.e., measure if training on generated data is able to make a model successful at testing on real settings. Our experiments compare different generators from the VAE and GAN framework on MNIST and fashion MNIST datasets. The results of our different experiments show that none of the generative models are able to replace completely true data to train a discriminative model. It also shows that the initial GAN and WGAN are the best choices to generate comparable datasets on MNIST and fashion MNIST but suffer from instability. VAE and CVAE are a bit less well-performing but are much more stable.

We propose a method of moments (MoM) algorithm for training large-scale implicit generative models. Moment estimation in this setting encounters two problems: it is often difficult to define the millions of moments needed to learn the model parameters, and it is hard to determine which properties are useful when specifying moments. To address the first issue, we introduce a moment network, and define the moments as the network's hidden units and the gradient of the network's output with the respect to its parameters. To tackle the second problem, we use asymptotic theory to highlight desiderata for moments -- namely they should minimize the asymptotic variance of estimated model parameters -- and introduce an objective to learn better moments. The sequence of objectives created by this Method of Learned Moments (MoLM) can train high-quality neural image samplers. On CIFAR-10, we demonstrate that MoLM-trained generators achieve significantly higher Inception Scores and lower Frechet Inception Distances than those trained with gradient penalty-regularized and spectrally-normalized adversarial objectives. These generators also achieve nearly perfect Multi-Scale Structural Similarity Scores on CelebA, and can create high-quality samples of 128x128 images.

The PC algorithm is a popular method for learning the structure of Gaussian Bayesian networks. It carries out statistical tests to determine absent edges in the network. It is hence governed by two parameters: (i) The type of test, and (ii) its significance level. These parameters are usually set to values recommended by an expert. Nevertheless, such an approach can suffer from human bias, leading to suboptimal reconstruction results. In this paper we consider a more principled approach for choosing these parameters in an automatic way. For this we optimize a reconstruction score evaluated on a set of different Gaussian Bayesian networks. This objective is expensive to evaluate and lacks a closed-form expression, which means that Bayesian optimization (BO) is a natural choice. BO methods use a model to guide the search and are hence able to exploit smoothness properties of the objective surface. We show that the parameters found by a BO method outperform those found by a random search strategy and the expert recommendation. Importantly, we have found that an often overlooked statistical test provides the best over-all reconstruction results.

Standard approaches to probabilistic reasoning require that one possesses an explicit model of the distribution in question. But, the empirical learning of models of probability distributions from partial observations is a problem for which efficient algorithms are generally not known. In this work we consider the use of bounded-degree fragments of the "sum-of-squares" logic as a probability logic. Prior work has shown that we can decide refutability for such fragments in polynomial-time. We propose to use such fragments to answer queries about whether a given probability distribution satisfies a given system of constraints and bounds on expected values. We show that in answering such queries, such constraints and bounds can be implicitly learned from partial observations in polynomial-time as well. It is known that this logic is capable of deriving many bounds that are useful in probabilistic analysis. We show here that it furthermore captures useful polynomial-time fragments of resolution. Thus, these fragments are also quite expressive.

One of the key challenges in applying reinforcement learning to real-life problems is that the amount of train-and-error required to learn a good policy increases drastically as the task becomes complex. One potential solution to this problem is to combine reinforcement learning with automated symbol planning and utilize prior knowledge on the domain. However, existing methods have limitations in their applicability and expressiveness. In this paper we propose a hierarchical reinforcement learning method based on abductive symbolic planning. The planner can deal with user-defined evaluation functions and is not based on the Herbrand theorem. Therefore it can utilize prior knowledge of the rewards and can work in a domain where the state space is unknown. We demonstrate empirically that our architecture significantly improves learning efficiency with respect to the amount of training examples on the evaluation domain, in which the state space is unknown and there exist multiple goals.

Long-lead forecasting for spatio-temporal problems can often entail complex nonlinear dynamics that are difficult to specify it a priori. Current statistical methodologies for modeling these processes are often overparameterized and thus, struggle from a computational perspective. One potential parsimonious solution to this problem is a method from the dynamical systems and engineering literature referred to as an echo state network (ESN). ESN models use so-called reservoir computing to efficiently estimate a dynamical neural network forecast, model referred to as a recurrent neural network (RNN). Moreover, so-called deep models have recently been shown to be successful at predicting high-dimensional complex nonlinear processes. These same traits can be used to characterize many spatio-temporal processes. Here we introduce a deep ensemble ESN (D-EESN) model. Through the use of an ensemble framework, this model is able to generate forecasts that are accompanied by uncertainty estimates. After introducing the D-EESN, we then develop a hierarchical Bayesian implementation. We use a general hierarchical Bayesian framework that accommodates non-Gaussian data types and multiple levels of uncertainties. The proposed methodology is first applied to a data set simulated from a novel non-Gaussian multiscale Lorenz-96 dynamical system simulation model and then to a long-lead United States (U.S.) soil moisture forecasting application.

Generative modeling of high dimensional data like images is a notoriously difficult and ill-defined problem. In particular, how to evaluate a learned generative model is unclear. In this position paper, we argue that adversarial learning, pioneered with generative adversarial networks (GANs), provides an interesting framework to implicitly define more meaningful task losses for generative modeling tasks, such as for generating "visually realistic" images. We refer to those task losses as parametric adversarial divergences and we give two main reasons why we think parametric divergences are good learning objectives for generative modeling. Additionally, we unify the processes of choosing a good structured loss (in structured prediction) and choosing a discriminator architecture (in generative modeling) using statistical decision theory; we are then able to formalize and quantify the intuition that "weaker" losses are easier to learn from, in a specific setting. Finally, we propose two new challenging tasks to evaluate parametric and nonparametric divergences: a qualitative task of generating very high-resolution digits, and a quantitative task of learning data that satisfies high-level algebraic constraints. We use two common divergences to train a generator and show that the parametric divergence outperforms the nonparametric divergence on both the qualitative and the quantitative task.