Deep reinforcement learning methods traditionally struggle with tasks where environment rewards are particularly sparse. One successful method of guiding exploration in these domains is to imitate trajectories provided by a human demonstrator. However, these demonstrations are typically collected under artificial conditions, i.e. with access to the agent's exact environment setup and the demonstrator's action and reward trajectories. Here we propose a two-stage method that overcomes these limitations by relying on noisy, unaligned footage without access to such data. First, we learn to map unaligned videos from multiple sources to a common representation using self-supervised objectives constructed over both time and modality (i.e. vision and sound). Second, we embed a single YouTube video in this representation to construct a reward function that encourages an agent to imitate human gameplay. This method of one-shot imitation allows our agent to convincingly exceed human-level performance on the infamously hard exploration games Montezuma's Revenge, Pitfall! and Private Eye for the first time, even if the agent is not presented with any environment rewards.

We show that when a third party, the adversary, steps into the two-party setting (agent and operator) of safely interruptible reinforcement learning, a trade-off has to be made between the probability of following the optimal policy in the limit, and the probability of escaping a dangerous situation created by the adversary. So far, the work on safely interruptible agents has assumed a perfect perception of the agent about its environment (no adversary), and therefore implicitly set the second probability to zero, by explicitly seeking a value of one for the first probability. We show that (1) agents can be made both interruptible and adversary-resilient, and (2) the interruptibility can be made safe in the sense that the agent itself will not seek to avoid it. We also solve the problem that arises when the agent does not go completely greedy, i.e. issues with safe exploration in the limit. Resilience to perturbed perception, safe exploration in the limit, and safe interruptibility are the three pillars of what we call \emph{virtuously safe reinforcement learning}.

We consider the transfer of experience samples (i.e., tuples < s, a, s', r >) in reinforcement learning (RL), collected from a set of source tasks to improve the learning process in a given target task. Most of the related approaches focus on selecting the most relevant source samples for solving the target task, but then all the transferred samples are used without considering anymore the discrepancies between the task models. In this paper, we propose a model-based technique that automatically estimates the relevance (importance weight) of each source sample for solving the target task. In the proposed approach, all the samples are transferred and used by a batch RL algorithm to solve the target task, but their contribution to the learning process is proportional to their importance weight. By extending the results for importance weighting provided in supervised learning literature, we develop a finite-sample analysis of the proposed batch RL algorithm. Furthermore, we empirically compare the proposed algorithm to state-of-the-art approaches, showing that it achieves better learning performance and is very robust to negative transfer, even when some source tasks are significantly different from the target task.

Inspired by how humans summarize long documents, we propose an accurate and fast summarization model that first selects salient sentences and then rewrites them abstractively (i.e., compresses and paraphrases) to generate a concise overall summary. We use a novel sentence-level policy gradient method to bridge the non-differentiable computation between these two neural networks in a hierarchical way, while maintaining language fluency. Empirically, we achieve the new state-of-the-art on all metrics (including human evaluation) on the CNN/Daily Mail dataset, as well as significantly higher abstractiveness scores. Moreover, by first operating at the sentence-level and then the word-level, we enable parallel decoding of our neural generative model that results in substantially faster (10-20x) inference speed as well as 4x faster training convergence than previous long-paragraph encoder-decoder models. We also demonstrate the generalization of our model on the test-only DUC-2002 dataset, where we achieve higher scores than a state-of-the-art model.

The reconstruction and analyzation of high energy particle physics data is just as important as the analyzation of the structure in real world networks. In a previous study it was explored how hierarchical clustering algorithms can be combined with kt cluster algorithms to provide a more generic clusterization method. Building on that, this paper explores the possibilities to involve deep learning in the process of cluster computation, by applying reinforcement learning techniques. The result is a model, that by learning on a modest dataset of 10; 000 nodes during 70 epochs can reach 83; 77% precision in predicting the appropriate clusters.

We present Value Propagation (VProp), a parameter-efficient differentiable planning module built on Value Iteration which can successfully be trained using reinforcement learning to solve unseen tasks, has the capability to generalize to larger map sizes, and can learn to navigate in dynamic environments. Furthermore, we show that the module enables learning to plan when the environment also includes stochastic elements, providing a cost-efficient learning system to build low-level size-invariant planners for a variety of interactive navigation problems. We evaluate on static and dynamic configurations of MazeBase grid-worlds, with randomly generated environments of several different sizes, and on a StarCraft navigation scenario, with more complex dynamics, and pixels as input.

Teaching agents to perform tasks using Reinforcement Learning is no easy feat. As the goal of reinforcement learning agents is to maximize the accumulated reward, they often find loopholes and misspecifications in the reward signal which lead to unwanted behavior. To overcome this, often, regularization is employed through the technique of reward shaping - the agent is provided an additional weighted reward signal, meant to lead it towards a desired behavior. The weight is considered as a hyper-parameter and is selected through trial and error, a time consuming and computationally intensive task. In this work, we present a novel multi-timescale approach for constrained policy optimization, called, 'Reward Constrained Policy Optimization' (RCPO), which enables policy regularization without the use of reward shaping. We prove the convergence of our approach and provide empirical evidence of its ability to train constraint satisfying policies.

Deep reinforcement learning (DRL) has proven to be an effective tool for creating general video-game AI. However most current DRL video-game agents learn end-to-end from the video-output of the game, which is superfluous for many applications and creates a number of additional problems. More importantly, directly working on pixel-based raw video data is substantially distinct from what a human player does.In this paper, we present a novel method which enables DRL agents to learn directly from object information. This is obtained via use of an object embedding network (OEN) that compresses a set of object feature vectors of different lengths into a single fixed-length unified feature vector representing the current game-state and fulfills the DRL simultaneously. We evaluate our OEN-based DRL agent by comparing to several state-of-the-art approaches on a selection of games from the GVG-AI Competition. Experimental results suggest that our object-based DRL agent yields performance comparable to that of those approaches used in our comparative study.

Policy gradient methods have been successfully applied to a variety of reinforcement learning tasks. However, while learning in a simulator, these methods do not utilise the opportunity to improve learning by adjusting certain environment variables: unobservable state features that are randomly determined by the environment in a physical setting, but that are controllable in a simulator. This can lead to slow learning, or convergence to highly suboptimal policies. In this paper, we present contextual policy optimisation (CPO). The central idea is to use Bayesian optimisation to actively select the distribution of the environment variable that maximises the improvement generated by each iteration of the policy gradient method. To make this Bayesian optimisation practical, we contribute two easy-to-compute low-dimensional fingerprints of the current policy. We apply CPO to a number of continuous control tasks of varying difficulty and show that CPO can efficiently learn policies that are robust to significant rare events, which are unlikely to be observable under random sampling but are key to learning good policies.

We present a study on reinforcement learning (RL) from human bandit feedback for sequence-to-sequence learning, exemplified by the task of bandit neural machine translation (NMT). We investigate the reliability of human bandit feedback, and analyze the influence of reliability on the learnability of a reward estimator, and the effect of the quality of reward estimates on the overall RL task. Our analysis of cardinal (5-point ratings) and ordinal (pairwise preferences) feedback shows that their intra- and inter-annotator $\alpha$-agreement is comparable. Best reliability is obtained for standardized cardinal feedback, and cardinal feedback is also easiest to learn and generalize from. Finally, improvements of over 1 BLEU can be obtained by integrating a regression-based reward estimator trained on cardinal feedback for 800 translations into RL for NMT. This shows that RL is possible even from small amounts of fairly reliable human feedback, pointing to a great potential for applications at larger scale.