We propose a method for offline learning of counterfactual predictions to address real world robotic reinforcement learning challenges. The proposed method encodes action-oriented visual observations as several "what if" questions learned offline from prior experience using reinforcement learning methods. These "what if" questions counterfactually predict how action-conditioned observation would evolve on multiple temporal scales if the agent were to stick to its current action. We show that combining these offline counterfactual predictions along with online in-situ observations (e.g. force feedback) allows efficient policy learning with only a sparse terminal (success/failure) reward. We argue that the learned predictions form an effective representation of the visual task, and guide the online exploration towards high-potential success interactions (e.g. contact-rich regions). Experiments were conducted in both simulation and real-world scenarios for evaluation. Our results demonstrate that it is practical to train a reinforcement learning agent to perform real-world fine manipulation in about half a day, without hand engineered perception systems or calibrated instrumentation. Recordings of the real robot training can be found via https://sites.google.com/view/realrl.

Sim-to-real, a term that describes where a model is trained in a simulator then transferred to the real world, is a technique that enables faster deep reinforcement learning (DRL) training. However, differences between the simulator and the real world often cause the model to perform poorly in the real world. Domain randomization is a way to bridge the sim-to-real gap by exposing the model to a wide range of scenarios so that it can generalize to real-world situations. However, following domain randomization to train an autonomous car racing model with DRL can lead to undesirable outcomes. Namely, a model trained with randomization tends to run slower; a higher completion rate on the testing track comes at the expense of longer lap times. This paper aims to boost the robustness of a trained race car model without compromising racing lap times. For a training track and a testing track having the same shape (and same optimal paths), but with different lighting, background, etc., we first train a model (teacher model) that overfits the training track, moving along a near optimal path. We then use this model to teach a student model the correct actions along with randomization. With our method, a model with 18.4\% completion rate on the testing track is able to help teach a student model with 52\% completion. Moreover, over an average of 50 trials, the student is able to finish a lap 0.23 seconds faster than the teacher. This 0.23 second gap is significant in tight races, with lap times of about 10 to 12 seconds.

Modifications to a neural network's input and output layers are often required to accommodate the specificities of most practical learning tasks. However, the impact of such changes on architecture's approximation capabilities is largely not understood. We present general conditions describing feature and readout maps that preserve an architecture's ability to approximate any continuous functions uniformly on compacts. As an application, we show that if an architecture is capable of universal approximation, then modifying its final layer to produce binary values creates a new architecture capable of deterministically approximating any classifier. In particular, we obtain guarantees for deep CNNs and deep feed-forward networks. Our results also have consequences within the scope of geometric deep learning. Specifically, when the input and output spaces are Cartan-Hadamard manifolds, we obtain geometrically meaningful feature and readout maps satisfying our criteria. Consequently, commonly used non-Euclidean regression models between spaces of symmetric positive definite matrices are extended to universal DNNs. The same result allows us to show that the hyperbolic feed-forward networks, used for hierarchical learning, are universal. Our result is also used to show that the common practice of randomizing all but the last two layers of a DNN produces a universal family of functions with probability one. We also provide conditions on a DNN's first (resp. last) few layer's connections and activation function which guarantee that these layers can have a width equal to the input (resp. output) space's dimension while not negatively affecting the architecture's approximation capabilities.

The neural attention mechanism plays an important role in many natural language processing applications. In particular, the use of multi-head attention extends single-head attention by allowing a model to jointly attend information from different perspectives. Without explicit constraining, however, multi-head attention may suffer from attention collapse, an issue that makes different heads extract similar attentive features, thus limiting the model's representation power. In this paper, for the first time, we provide a novel understanding of multi-head attention from a Bayesian perspective. Based on the recently developed particle-optimization sampling techniques, we propose a non-parametric approach that explicitly improves the repulsiveness in multi-head attention and consequently strengthens model's expressiveness. Remarkably, our Bayesian interpretation provides theoretical inspirations on the not-well-understood questions: why and how one uses multi-head attention. Extensive experiments on various attention models and applications demonstrate that the proposed repulsive attention can improve the learned feature diversity, leading to more informative representations with consistent performance improvement on various tasks.

Policy iteration (PI) is a recursive process of policy evaluation and improvement for solving an optimal decision-making/control problem, or in other words, a reinforcement learning (RL) problem. PI has also served as the fundamental for developing RL methods. In this paper, we propose two PI methods, called differential PI (DPI) and integral PI (IPI), and their variants, for a general RL framework in continuous time and space (CTS), where the environment is modeled by a system of ordinary differential equations (ODEs). The proposed methods inherit the current ideas of PI in classical RL and optimal control and theoretically support the existing RL algorithms in CTS: TD-learning and value-gradient-based (VGB) greedy policy update. We also provide case studies including 1) discounted RL and 2) optimal control tasks. Fundamental mathematical properties -- admissibility, uniqueness of the solution to the Bellman equation (BE), monotone improvement, convergence, and optimality of the solution to the Hamilton-Jacobi-Bellman equation (HJBE) -- are all investigated in-depth and improved from the existing theory, along with the general and case studies. Finally, the proposed ones are simulated with an inverted-pendulum model and their model-based and partially model-free implementations to support the theory and further investigate them beyond.

Despite empirical success, the theory of reinforcement learning (RL) with value function approximation remains fundamentally incomplete. Prior work has identified a variety of pathological behaviours that arise in RL algorithms that combine approximate on-policy evaluation and greedification. One prominent example is policy oscillation, wherein an algorithm may cycle indefinitely between policies, rather than converging to a fixed point. What is not well understood however is the quality of the policies in the region of oscillation. In this paper we present simple examples illustrating that in addition to policy oscillation and multiple fixed points -- the same basic issue can lead to convergence to the worst possible policy for a given approximation. Such behaviours can arise when algorithms optimize evaluation accuracy weighted by the distribution of states that occur under the current policy, but greedify based on the value of states which are rare or nonexistent under this distribution. This means the values used for greedification are unreliable and can steer the policy in undesirable directions. Our observation that this can lead to the worst possible policy shows that in a general sense such algorithms are unreliable. The existence of such examples helps to narrow the kind of theoretical guarantees that are possible and the kind of algorithmic ideas that are likely to be helpful. We demonstrate analytically and experimentally that such pathological behaviours can impact a wide range of RL and dynamic programming algorithms; such behaviours can arise both with and without bootstrapping, and with linear function approximation as well as with more complex parameterized functions like neural networks.

Query-based document summarization aims to extract or generate a summary of a document which directly answers or is relevant to the search query. It is an important technique that can be beneficial to a variety of applications such as search engines, document-level machine reading comprehension, and chatbots. Currently, datasets designed for query-based summarization are short in numbers and existing datasets are also limited in both scale and quality. Moreover, to the best of our knowledge, there is no publicly available dataset for Chinese query-based document summarization. In this paper, we present QBSUM, a high-quality large-scale dataset consisting of 49,000+ data samples for the task of Chinese query-based document summarization. We also propose multiple unsupervised and supervised solutions to the task and demonstrate their high-speed inference and superior performance via both offline experiments and online A/B tests. The QBSUM dataset is released in order to facilitate future advancement of this research field.

We present an information-based uncertainty quantification method for general Markov Random Fields. Markov Random Fields (MRFs) are structured, probabilistic graphical models over undirected graphs, and provide a fundamental unifying modeling tool for statistical mechanics, probabilistic machine learning, and artificial intelligence. Typically MRFs are complex and high-dimensional with nodes and edges (connections) built in a modular fashion from simpler, low-dimensional probabilistic models and their local connections; in turn, this modularity allows to incorporate available data to MRFs and efficiently simulate them by leveraging their graph-theoretic structure. Learning graphical models from data and/or constructing them from physical modeling and constraints necessarily involves uncertainties inherited from data, modeling choices, or numerical approximations. These uncertainties in the MRF can be manifested either in the graph structure or the probability distribution functions, and necessarily will propagate in predictions for quantities of interest. Here we quantify such uncertainties using tight, information-based bounds on the predictions of quantities of interest; these bounds take advantage of the graphical structure of MRFs and are capable of handling the inherent high-dimensionality of such graphical models. We demonstrate our methods in MRFs for medical diagnostics and statistical mechanics models. In the latter, we develop uncertainty quantification bounds for finite-size effects and phase diagrams, which constitute two of the typical predictions goals of statistical mechanics modeling.

Rising penetration levels of (residential) photovoltaic (PV) power as distributed energy resource pose a number of challenges to the electricity infrastructure. High quality, general tools to provide accurate forecasts of power production are urgently needed. In this article, we propose a supervised deep learning model for end-to-end forecasting of PV power production. The proposed model is based on two seminal concepts that led to significant performance improvements of deep learning approaches in other sequence-related fields, but not yet in the area of time series prediction: the sequence to sequence architecture and attention mechanism as a context generator. The proposed model leverages numerical weather predictions and high-resolution historical measurements to forecast a binned probability distribution over the prognostic time intervals, rather than the expected values of the prognostic variable. This design offers significant performance improvements compared to common baseline approaches, such as fully connected neural networks and one-block long short-term memory architectures. Using normalized root mean square error based forecast skill score as a performance indicator, the proposed approach is compared to other models. The results show that the new design performs at or above the current state of the art of PV power forecasting.

We consider the problem of local planning in fixed-horizon Markov Decision Processes (MDPs) with linear function approximation and a generative model under the assumption that the optimal action-value function lies in the span of a feature map that is available to the planner. Previous work has left open the question of whether there exists sound planners that need only poly(H, d) queries regardless of the MDP, where H is the horizon and d is the dimensionality of the features. We answer this question in the negative: we show that any sound planner must query at least min(exp({\Omega}(d)), {\Omega}(2^H)) samples. We also show that for any {\delta}>0, the least-squares value iteration algorithm with O(H^5d^(H+1)/{\delta}^2) queries can compute a {\delta}-optimal policy. We discuss implications and remaining open questions.