Imitation learning seeks to learn an expert policy from sampled demonstrations. However, in the real world, it is often difficult to find a perfect expert and avoiding dangerous behaviors becomes relevant for safety reasons. We present the idea of \textit{learning to avoid}, an objective opposite to imitation learning in some sense, where an agent learns to avoid a demonstrator policy given an environment. We define avoidance learning as the process of optimizing the agent's reward while avoiding dangerous behaviors given by a demonstrator. In this work we develop a framework of avoidance learning by defining a suitable objective function for these problems which involves the \emph{distance} of state occupancy distributions of the expert and demonstrator policies. We use density estimates for state occupancy measures and use the aforementioned distance as the reward bonus for avoiding the demonstrator. We validate our theory with experiments using a wide range of partially observable environments. Experimental results show that we are able to improve sample efficiency during training compared to state of the art policy optimization and safety methods.
In tabular case, when the reward and environment dynamics are known, policy evaluation can be written as $\bm{V}_{\bm{\pi}} = (I - \gamma P_{\bm{\pi}})^{-1} \bm{r}_{\bm{\pi}}$, where $P_{\bm{\pi}}$ is the state transition matrix given policy ${\bm{\pi}}$ and $\bm{r}_{\bm{\pi}}$ is the reward signal given ${\bm{\pi}}$. What annoys us is that $P_{\bm{\pi}}$ and $\bm{r}_{\bm{\pi}}$ are both mixed with ${\bm{\pi}}$, which means every time when we update ${\bm{\pi}}$, they will change together. In this paper, we leverage the notation from \cite{wang2007dual} to disentangle ${\bm{\pi}}$ and environment dynamics which makes optimization over policy more straightforward. We show that policy gradient theorem \cite{sutton2018reinforcement} and TRPO \cite{schulman2015trust} can be put into a more general framework and such notation has good potential to be extended to model-based reinforcement learning.
Batch normalization has been widely used to improve optimization in deep neural networks. While the uncertainty in batch statistics can act as a regularizer, using these dataset statistics specific to the training set impairs generalization in certain tasks. Recently, alternative methods for normalizing feature activations in neural networks have been proposed. Among them, group normalization has been shown to yield similar, in some domains even superior performance to batch normalization. All these methods utilize a learned affine transformation after the normalization operation to increase representational power. Methods used in conditional computation define the parameters of these transformations as learnable functions of conditioning information. In this work, we study whether and where the conditional formulation of group normalization can improve generalization compared to conditional batch normalization. We evaluate performances on the tasks of visual question answering, few-shot learning, and conditional image generation.
Goal-conditioned policies are used in order to break down complex reinforcement learning (RL) problems by using subgoals, which can be defined either in state space or in a latent feature space. This can increase the efficiency of learning by using a curriculum, and also enables simultaneous learning and generalization across goals. A crucial requirement of goal-conditioned policies is to be able to determine whether the goal has been achieved. Having a notion of distance to a goal is thus a crucial component of this approach. However, it is not straightforward to come up with an appropriate distance, and in some tasks, the goal space may not even be known a priori. In this work we learn a distance-to-goal estimate which is computed in terms of the number of actions that would need to be carried out in a self-supervised approach. Our method solves complex tasks without prior domain knowledge in the online setting in three different scenarios in the context of goal-conditioned policies a) the goal space is the same as the state space b) the goal space is given but an appropriate distance is unknown and c) the state space is accessible, but only a subset of the state space represents desired goals, and this subset is known a priori. We also propose a goal-generation mechanism as a secondary contribution.
Despite continuing medical advances, the rate of newborn morbidity and mortality globally remains high, with over 6 million casualties every year. The prediction of pathologies affecting newborns based on their cry is thus of significant clinical interest, as it would facilitate the development of accessible, low-cost diagnostic tools\cut{ based on wearables and smartphones}. However, the inadequacy of clinically annotated datasets of infant cries limits progress on this task. This study explores a neural transfer learning approach to developing accurate and robust models for identifying infants that have suffered from perinatal asphyxia. In particular, we explore the hypothesis that representations learned from adult speech could inform and improve performance of models developed on infant speech. Our experiments show that models based on such representation transfer are resilient to different types and degrees of noise, as well as to signal loss in time and frequency domains.
Recently, neural network based approaches have achieved significant improvement for solving large, complex, graph-structured problems. However, their bottlenecks ] still need to be addressed, and the advantages of multi-scale information and deep architectures have not been sufficiently exploited. In this paper, we theoretically analyze how existing Graph Convolutional Networks (GCNs) have limited expressive power due to the constraint of the activation functions and their architectures. We generalize spectral graph convolution and deep GCN in block Krylov subspace forms and devise two architectures, both with the potential to be scaled deeper but each making use of the multi-scale information in different ways. We further show that the equivalence of these two architectures can be established under certain conditions. On several node classification tasks, with or without the help of validation, the two new architectures achieve better performance compared to many state-of-the-art methods.
Stochastic variance-reduced gradient (SVRG) is an optimization method originally designed for tackling machine learning problems with a finite sum structure. SVRG was later shown to work for policy evaluation, a problem in reinforcement learning in which one aims to estimate the value function of a given policy. SVRG makes use of gradient estimates at two scales. At the slower scale, SVRG computes a full gradient over the whole dataset, which could lead to prohibitive computation costs. In this work, we show that two variants of SVRG for policy evaluation could significantly diminish the number of gradient calculations while preserving a linear convergence speed. More importantly, our theoretical result implies that one does not need to use the entire dataset in every epoch of SVRG when it is applied to policy evaluation with linear function approximation. Our experiments demonstrate large computational savings provided by the proposed methods.
To improve the speed and accuracy of the trace based policy evaluation method TD({\lambda}), under appropriate assumptions, we derive and propose an off-policy compatible method of meta-learning state-based {\lambda}'s online with efficient incremental updates. Furthermore, we prove the derived bias-variance tradeoff minimization method, with slight adjustments, is equivalent to minimizing the overall target error in terms of state based {\lambda}'s. In experiments, the method shows significantly better performance when compared to the existing method and the baselines.
Despite recent successes in Reinforcement Learning, value-based methods often suffer from high variance hindering performance. In this paper, we illustrate this in a continuous control setting where state of the art methods perform poorly whenever sensor noise is introduced. To overcome this issue, we introduce Recurrent Value Functions (RVFs) as an alternative to estimate the value function of a state. We propose to estimate the value function of the current state using the value function of past states visited along the trajectory. Due to the nature of their formulation, RVFs have a natural way of learning an emphasis function that selectively emphasizes important states. First, we establish RVF's asymptotic convergence properties in tabular settings. We then demonstrate their robustness on a partially observable domain and continuous control tasks. Finally, we provide a qualitative interpretation of the learned emphasis function.