This work presents an exploration and imitation-learning-based agent capable of state-of-the-art performance in playing text-based computer games. Text-based computer games describe their world to the player through natural language and expect the player to interact with the game using text. These games are of interest as they can be seen as a testbed for language understanding, problem-solving, and language generation by artificial agents. Moreover, they provide a learning environment in which these skills can be acquired through interactions with an environment rather than using fixed corpora. One aspect that makes these games particularly challenging for learning agents is the combinatorially large action space. Existing methods for solving text-based games are limited to games that are either very simple or have an action space restricted to a predetermined set of admissible actions. In this work, we propose to use the exploration approach of Go-Explore for solving text-based games. More specifically, in an initial exploration phase, we first extract trajectories with high rewards, after which we train a policy to solve the game by imitating these trajectories. Our experiments show that this approach outperforms existing solutions in solving text-based games, and it is more sample efficient in terms of the number of interactions with the environment. Moreover, we show that the learned policy can generalize better than existing solutions to unseen games without using any restriction on the action space.

Interactive Fiction games are text-based simulations in which an agent interacts with the world purely through natural language. They are ideal environments for studying how to extend reinforcement learning agents to meet the challenges of natural language understanding, partial observability, and action generation in combinatorially-large text-based action spaces. We present KG-A2C, an agent that builds a dynamic knowledge graph while exploring and generates actions using a template-based action space. We contend that the dual uses of the knowledge graph to reason about game state and to constrain natural language generation are the keys to scalable exploration of combinatorially large natural language actions. Results across a wide variety of IF games show that KG-A2C outperforms current IF agents despite the exponential increase in action space size.

Generating natural questions from an image is a semantic task that requires using vision and language modalities to learn multimodal representations. Images can have multiple visual and language cues such as places, captions, and tags. In this paper, we propose a principled deep Bayesian learning framework that combines these cues to produce natural questions. We observe that with the addition of more cues and by minimizing uncertainty in the among cues, the Bayesian network becomes more confident. We propose a Minimizing Uncertainty of Mixture of Cues (MUMC), that minimizes uncertainty present in a mixture of cues experts for generating probabilistic questions. This is a Bayesian framework and the results show a remarkable similarity to natural questions as validated by a human study. We observe that with the addition of more cues and by minimizing uncertainty among the cues, the Bayesian framework becomes more confident. Ablation studies of our model indicate that a subset of cues is inferior at this task and hence the principled fusion of cues is preferred. Further, we observe that the proposed approach substantially improves over state-of-the-art benchmarks on the quantitative metrics (BLEU-n, METEOR, ROUGE, and CIDEr). Here we provide project link for Deep Bayesian VQG \url{https://delta-lab-iitk.github.io/BVQG/}

Relational autocompletion is the problem of automatically filling out some missing fields in a relational database. We tackle this problem within the probabilistic logic programming framework of Distributional Clauses (DC), which supports both discrete and continuous probability distributions. Within this framework, we introduce Dreaml -- an approach to learn both the structure and the parameters of DC programs from databases that may contain missing information. To realize this, Dreaml integrates statistical modeling, distributional clauses with rule learning. The distinguishing features of Dreaml are that it 1) tackles relational autocompletion, 2) learns distributional clauses extended with statistical models, 3) deals with both discrete and continuous distributions, 4) can exploit background knowledge, and 5) uses an expectation-maximization based algorithm to cope with missing data.

As social beings, much human behavior is predicated on social context - the ambient social state that includes cultural norms, social signals, individual preferences, etc. In this paper, we propose a socially-aware task and motion planning algorithm that considers social context to generate appropriate and effective plans in human social environments (HSEs). The key strength of our proposed approach is that it explicitly models how potential actions not only affect objective cost, but also transform the social context in which it plans and acts. We investigate strategies to limit the complexity of our algorithm, so that our planner will remain tractable for mobile platforms in complex HSEs like hospitals and factories. The planner will also consider the relative importance and urgency of its tasks, which it uses to determine when it is and is not appropriate to violate social expectations to achieve its objective. This social awareness will allow robots to understand a fundamental rule of society: just because something makes your job easier, does not make it the right thing to do! To our knowledge, the proposed work is the first task and motion planning approach that supports socially intelligent robot policy for HSEs. Through this ongoing work, robots will be able to understand, respect, and leverage social context accomplish tasks both acceptably and effectively in HSEs.

We give a simple optimistic algorithm for which it is easy to derive regret bounds of O(sqrt{t-mix SAT}) steps in uniformly ergodic Markov decision processes with S states, A actions, and mixing time parameter t-mix. These bounds are the first regret bounds in the general, non-episodic setting with an optimal dependence on all given parameters. They could only be improved by using an alternative mixing time parameter.

What initially got you interested in artificial intelligence? I was reading a book on probability when I came across a famous theorem. At the time, I naively wondered if I could apply this theorem into a natural language problem I was attempting to solve at work. As it turns out, the algorithm already existed unbeknownst to me, it was called the Naïve Bayes, a very famous and simple generative model used in classical machine learning. That theorem was Bayes theorem.

In bipartite networks, community structures are restricted to being disassortative, in that nodes of one type are grouped according to common patterns of connection with nodes of the other type. This makes the stochastic block model (SBM), a highly flexible generative model for networks with block structure, an intuitive choice for bipartite community detection. However, typical formulations of the SBM do not make use of the special structure of bipartite networks. In this work, we introduce a Bayesian nonparametric formulation of the SBM and a corresponding algorithm to efficiently find communities in bipartite networks without overfitting. The biSBM improves community detection results over general SBMs when data are noisy, improves the model resolution limit by a factor of $\sqrt{2}$, and expands our understanding of the complicated optimization landscape associated with community detection tasks. A direct comparison of certain terms of the prior distributions in the biSBM and a related high-resolution hierarchical SBM also reveals a counterintuitive regime of community detection problems, populated by smaller and sparser networks, where non-hierarchical models outperform their more flexible counterpart.

Throughout the course of the paper, we assume that y and X have been centered at 0 so there is no intercept in the model, where y is the n 1 vector of centered responses, X is the n p matrix of standardized regressors, β is the p 1 vector of coefficients to be estimated, and null is the n 1 vector of independent and identically distributed normal errors with mean 0 and variance σ 2 . Compared to traditional penalization functions that are usually symmetric about 0, continuous and nondecreasing in (0,), the rLASSO penalty functions are decreasing in (0,), discontinuous at 0, and converge to infinity when the coefficients approach zero. From a theoretical standpoint, rLASSO shares the same oracle property and same rate of estimation error with other LASSOtype penalty functions. An early reference to this class of models can be found in Song and Liang (2015), with more recent papers focusing on large sample asymptotics, along with computational strategies for frequentist estimation (Shin et al., 2018; Song, 2018). Our approach differs from this line of work in adopting a Bayesian perspective on rLASSO estimation. Ideally, a Bayesian solution can be obtained by placing appropriate priors on the regression coefficients that will mimic the effects of the rLASSO penalty. As apparent from (1), this arises in assuming a prior for β that decomposes as a product of independent inverse Laplace (double exponential) densities: π (β) p null j 1 λ 2β 2 j exp{ λ β j }I { β j null 0 }.

Uncertainty quantification for deep learning is a challenging open problem. Bayesian statistics offer a mathematically grounded framework to reason about uncertainties; however, approximate posteriors for modern neural networks still require prohibitive computational costs. We propose a family of algorithms which split the classification task into two stages: representation learning and uncertainty estimation. We compare four specific instances, where uncertainty estimation is performed via either an ensemble of Stochastic Gradient Descent or Stochastic Gradient Langevin Dynamics snapshots, an ensemble of bootstrapped logistic regressions, or via a number of Monte Carlo Dropout passes. We evaluate their performance in terms of \emph{selective} classification (risk-coverage), and their ability to detect out-of-distribution samples. Our experiments suggest there is limited value in adding multiple uncertainty layers to deep classifiers, and we observe that these simple methods strongly outperform a vanilla point-estimate SGD in some complex benchmarks like ImageNet.