This paper investigates how to efficiently transition and update policies, trained initially with demonstrations, using off-policy actor-critic reinforcement learning. It is well-known that techniques based on Learning from Demonstrations, for example behavior cloning, can lead to proficient policies given limited data. However, it is currently unclear how to efficiently update that policy using reinforcement learning as these approaches are inherently optimizing different objective functions. Previous works have used loss functions which combine behavioral cloning losses with reinforcement learning losses to enable this update, however, the components of these loss functions are often set anecdotally, and their individual contributions are not well understood. In this work we propose the Cycle-of-Learning (CoL) framework that uses an actor-critic architecture with a loss function that combines behavior cloning and 1-step Q-learning losses with an off-policy pre-training step from human demonstrations. This enables transition from behavior cloning to reinforcement learning without performance degradation and improves reinforcement learning in terms of overall performance and training time. Additionally, we carefully study the composition of these combined losses and their impact on overall policy learning. We show that our approach outperforms state-of-the-art techniques for combining behavior cloning and reinforcement learning for both dense and sparse reward scenarios. Our results also suggest that directly including the behavior cloning loss on demonstration data helps to ensure stable learning and ground future policy updates.

Jointly quantifying the contribution of the learning algorithm and training dataGal Yona Amirata Ghorbani James Zou Weizmann Institute Stanford University Stanford University Abstract A fancy learning algorithm A outperforms a baseline method B when they are both trained on the same data. Should A get all of the credit for the improved performance or does the training data also deserve some credit? When deployed in a new setting from a different domain, however, A makes more mistakes than B . How much of the blame should go to the learning algorithm or the training data? Such questions are becoming increasingly important and prevalent as we aim to make ML more accountable. Their answers would also help us allocate resources between algorithm design and data collection. In this paper, we formalize these questions and provide a principled Extended Shapley framework to jointly quantify the contribution of the learning algorithm and training data. Extended Shapley uniquely satisfies several natural properties that ensure equitable treatment of data and algorithm. Through experiments and theoretical analysis, we demonstrate that Extended Shapley has several important applications: 1) it provides a new metric of ML performance improvement that disentangles the influence of the data regime and the algorithm; 2) it facilitates ML accountability by properly assigning responsibility for mistakes; 3) it provides more robustness to manipulation by the ML designer. Introduction In machine learning (ML), the standard way to evaluate a new learning algorithm A is to compare its performance with the performance of a baseline algorithm B, when A and B are trained on the same dataset D . For example, if A and B achieves 0.9 and 0.7 accuracy, then papers typically report that A is better than B by 0.2. Implicit in this ubiquitous practice is the assumption that A itself is solely responsible for all of the difference in performance.

The goal of representation learning of knowledge graph is to encode both entities and relations into a low-dimensional embedding spaces. Many recent works have demonstrated the benefits of knowledge graph embedding on knowledge graph completion task, such as relation extraction. However, we observe that: 1) existing method just take direct relations between entities into consideration and fails to express high-order structural relationship between entities; 2) these methods just leverage relation triples of KGs while ignoring a large number of attribute triples that encoding rich semantic information. To overcome these limitations, this paper propose a novel knowledge graph embedding method, named KANE, which is inspired by the recent developments of graph convolutional networks (GCN). KANE can capture both high-order structural and attribute information of KGs in an efficient, explicit and unified manner under the graph convolutional networks framework. Empirical results on three datasets show that KANE significantly outperforms seven state-of-arts methods. Further analysis verify the efficiency of our method and the benefits brought by the attention mechanism.

A BSTRACT V alue-based reinforcement learning (RL) methods like Q-learning have shown success in a variety of domains. One challenge in applying Q-learning to continuous-action RL problems, however, is the continuous action maximization ( max-Q) required for optimal Bellman backup. In this work, we develop CAQL, a (class of) algorithm(s) for continuous-action Q-learning that can use several plug-and-play optimizers for the max-Q problem. Leveraging recent optimization results for deep neural networks, we show that max-Q can be solved optimally using mixed-integer programming (MIP) . When the Q-function representation has sufficient power, MIP-based optimization gives rise to better policies and is more robust than approximate methods (e.g., gradient ascent, cross-entropy search). We further develop several techniques to accelerate inference in CAQL, which despite their approximate nature, perform well. We compare CAQL with state-of-the-art RL algorithms on benchmark continuous-control problems that have different degrees of action constraints and show that CAQL outperforms policy-based methods in heavily constrained environments, often dramatically. When the action space is finite, value-based algorithms such as Q-learning (Watkins & Dayan, 1992), which implicitly finds a policy by learning the optimal value function, are often very efficient because action optimization can be done by exhaustive enumeration. By contrast, in problems with a continuous action spaces (e.g., robotics (Peters & Schaal, 2006)), policy-based algorithms, such as policy gradient (PG) (Sutton et al., 2000; Silver et al., 2014) or cross-entropy policy search (CEPS) (Mannor et al., 2003; Kalashnikov et al., 2018), which directly learn a return-maximizing policy, have proven more practical. Recently, methods such as ensemble critic (Fujimoto et al., 2018) and entropy regularization (Haarnoja et al., 2018) have been developed to improve the performance of policy-based RL algorithms. Policy-based approaches require a reasonable choice of policy parameterization. In some continuous control problems, Gaussian distributions over actions conditioned on some state representation is used. However, in applications such as RSs, where actions often take the form of high-dimensional item-feature vectors, policies cannot typically be modeled by common action distributions.

Humans are masters at quickly learning many complex tasks, relying on an approximate understanding of the dynamics of their environments. In much the same way, we would like our learning agents to quickly adapt to new tasks. In this paper, we explore how model-based Reinforcement Learning (RL) can facilitate transfer to new tasks. We develop an algorithm that learns an action-conditional, predictive model of expected future observations, rewards and values from which a policy can be derived by following the gradient of the estimated value along imagined trajectories. We show how robust policy optimization can be achieved in robot manipulation tasks even with approximate models that are learned directly from vision and proprioception. We evaluate the efficacy of our approach in a transfer learning scenario, re-using previously learned models on tasks with different reward structures and visual distractors, and show a significant improvement in learning speed compared to strong off-policy baselines. Videos with results can be found at https://sites.google.com/view/ivg-corl19

A BSTRACT Biological evolution has distilled the experiences of many learners into the general learning algorithms of humans. Our novel meta-reinforcement learning algorithm MetaGenRL is inspired by this process. MetaGenRL distills the experiences of many complex agents to meta-learn a low-complexity neural objective function that affects how future individuals will learn. Unlike recent meta-RL algorithms, MetaGenRL can generalize to new environments that are entirely different from those used for meta-training. In some cases, it even outperforms human-engineered RL algorithms. MetaGenRL uses off-policy second-order gradients during meta-training that greatly increase its sample efficiency. 1 I NTRODUCTION The process of evolution has equipped humans with incredibly general learning algorithms. They allow us to flexibly solve a wide range of problems, even in the absence of many related prior experiences. The inductive biases that give rise to these capabilities are the result of distilling the collective experiences of many learners throughout the course of natural evolution. By essentially learning from learning experiences in this way, this knowledge can be compactly encoded in the genetic code of an individual to give rise to the general learning capabilities that we observe today. In contrast, Reinforcement Learning (RL) in artificial agents rarely proceeds in this way. The learning rules that are used to train agents are the result of years of human engineering and design, (e.g. Correspondingly, artificial agents are inherently limited by the ability of the designer to incorporate the right inductive biases in order to learn from previous experiences.

Leveraging an equivalence property in the state-space of a Markov Decision Process (MDP) has been investigated in several studies. This paper studies equivalence structure in the reinforcement learning (RL) setup, where transition distributions are no longer assumed to be known. We present a notion of similarity between transition probabilities of various state-action pairs of an MDP, which naturally defines an equivalence structure in the state-action space. We present equivalence-aware confidence sets for the case where the learner knows the underlying structure in advance. These sets are provably smaller than their corresponding equivalence-oblivious counterparts. In the more challenging case of an unknown equivalence structure, we present an algorithm called ApproxEquivalence that seeks to find an (approximate) equivalence structure, and define confidence sets using the approximate equivalence. To illustrate the efficacy of the presented confidence sets, we present C-UCRL, as a natural modification of UCRL2 for RL in undiscounted MDPs. In the case of a known equivalence structure, we show that C-UCRL improves over UCRL2 in terms of regret by a factor of $\sqrt{SA/C}$, in any communicating MDP with $S$ states, $A$ actions, and $C$ classes, which corresponds to a massive improvement when $C \ll SA$. To the best of our knowledge, this is the first work providing regret bounds for RL when an equivalence structure in the MDP is efficiently exploited. In the case of an unknown equivalence structure, we show through numerical experiments that C-UCRL combined with ApproxEquivalence outperforms UCRL2 in ergodic MDPs.

Over its lifetime, a reinforcement learning agent is often tasked with different tasks. How to efficiently adapt a previously learned control policy from one task to another, remains an open research question. In this paper, we investigate how instructions formulated in natural language can enable faster and more effective task adaptation. This can serve as the basis for developing language instructed skills, which can be used in a lifelong learning setting. Our method is capable of assessing, given a set of developed base control policies, which policy will adapt best to a new unseen task.

We present a first attempt to elucidate an Information-Theoretic approach to design the reward provided by a natural language environment to some structure learning agent. To this end, we revisit the Information Theory of unsupervised induction of phrase-structure grammars to characterize the behavior of simulated agents whose actions are characterized in terms of random sets of linguistic samples. Our results showed empirical evidence of that semantic structures (built using Open Information Extraction methods) can be distinguished from randomly constructed structures by observing the Mutual Information among their constituent linguistic random sets. This suggests the possibility of rewarding structure learning agents without using pretrained structural analyzers (oracle actors or experts).

-- This work addresses the problem of semantic scene understanding under foggy road conditions. Although marked progress has been made in semantic scene understanding over the recent years, it is mainly concentrated on clear weather outdoor scenes. Extending semantic segmentation methods to adverse weather conditions like fog is crucially important for outdoor applications such as self-driving cars. In this paper, we propose a novel method, which uses purely synthetic data to improve the performance on unseen real-world foggy scenes captured in the streets of Zurich and its surroundings. Our results highlight the potential and power of photo-realistic synthetic images for training and especially fine-tuning deep neural nets. Our contributions are threefold, 1) we created a purely synthetic, high-quality foggy dataset of 25,000 unique outdoor scenes, that we call Foggy Synscapes and plan to release publicly 2) we show that with this data we outperform previous approaches on real-world foggy test data 3) we show that a combination of our data and previously used data can even further improve the performance on real-world foggy data. The last years have seen tremendous progress in tasks relevant to autonomous driving [1].