How to best explore in domains with sparse, delayed, and deceptive rewards is an important open problem for reinforcement learning (RL). This paper considers one such domain, the recently-proposed multi-agent benchmark of Pommerman. This domain is very challenging for RL --- past work has shown that model-free RL algorithms fail to achieve significant learning without artificially reducing the environment's complexity. In this paper, we illuminate reasons behind this failure by providing a thorough analysis on the hardness of random exploration in Pommerman. While model-free random exploration is typically futile, we develop a model-based automatic reasoning module that can be used for safer exploration by pruning actions that will surely lead the agent to death. We empirically demonstrate that this module can significantly improve learning.

Deep reinforcement learning has achieved great successes in recent years, however, one main challenge is the sample inefficiency. In this paper, we focus on how to use action guidance by means of a non-expert demonstrator to improve sample efficiency in a domain with sparse, delayed, and possibly deceptive rewards: the recently-proposed multi-agent benchmark of Pommerman. We propose a new framework where even a non-expert simulated demonstrator, e.g., planning algorithms such as Monte Carlo tree search with a small number rollouts, can be integrated within asynchronous distributed deep reinforcement learning methods. Compared to a vanilla deep RL algorithm, our proposed methods both learn faster and converge to better policies on a two-player mini version of the Pommerman game. Introduction Deep reinforcement learning (DRL) has enabled better scalability and generalization for challenging domains (Arulku-maran et al. 2017; Li 2017; Hernandez-Leal, Kartal, and Taylor 2018) such as Atari games (Mnih et al. 2015), Go (Silver et al. 2016) and multiagent games (e.g., Starcraft II and DOT A 2) (OpenAI 2018). However, one of the current biggest challenges for DRL is sample efficiency (Y u 2018). On the one hand, once a DRL agent is trained, it can be deployed to act in real-time by only performing an inference through the trained model. On the other hand, planning methods such as Monte Carlo tree search (MCTS) (Browne et al. 2012) do not have a training phase, but they perform computationally costly simulation based rollouts (assuming access to a simulator) to find the best action to take. There are several ways to get the best of both DRL and search methods.

Recent progress in the field of reinforcement learning has been accelerated by virtual learning environments such as video games, where novel algorithms and ideas can be quickly tested in a safe and reproducible manner. We introduce the Google Research F ootball Environment, a new reinforcement learning environment where agents are trained to play football in an advanced, physics-based 3D simulator. The resulting environment is challenging, easy to use and customize, and it is available under a permissive open-source license. In addition, it provides support for multiplayer and multi-agent experiments. We propose three full-game scenarios of varying difficulty with the F ootball Benchmarks and report baseline results for three commonly used reinforcement algorithms (IMP ALA, PPO, and Ape-X DQN). We also provide a diverse set of simpler scenarios with the F ootball Academy and showcase several promising research directions.

In many real-world decision making problems, reaching an optimal decision requires taking into account a variable number of objects around the agent. Autonomous driving is a domain in which this is especially relevant, since the number of cars surrounding the agent varies considerably over time and affects the optimal action to be taken. Classical methods that process object lists can deal with this requirement. However, to take advantage of recent high-performing methods based on deep reinforcement learning in modular pipelines, special architectures are necessary. For these, a number of options exist, but a thorough comparison of the different possibilities is missing. In this paper, we elaborate limitations of fully-connected neural networks and other established approaches like convolutional and recurrent neural networks in the context of reinforcement learning problems that have to deal with variable sized inputs. We employ the structure of Deep Sets in off-policy reinforcement learning for high-level decision making, highlight their capabilities to alleviate these limitations, and show that Deep Sets not only yield the best overall performance but also offer better generalization to unseen situations than the other approaches.

Lung sounds auscultation is the first and most common examination carried out by every general practitioner or family doctor. It is fast, easy and well known procedure, popularized by La ennec (Hy-acinthe, 1819), who invented the stethoscope. Nowadays, different variants of such tool can be found on the market, both analog and electronic, but regardless of the type of stethoscope, this process still is highly subjective. Indeed, an auscultation normally involves the usage of a stethoscope by a physician, thus relying on the examiner's own hearing, experience and ability to interpret psychoacoustical features. Another strong limitation of standard auscultation can be found in the stethoscope itself, since its frequency response tends to attenuate frequency components of the lung sound signal above nearly 120 Hz, leaving lower frequency bands to be analyzed and to which the human ear is not really sensitive (Sovijrvi et al., 2000) (Sarkar et al., 2015).

Many animals, and an increasing number of artificial agents, display sophisticated capabilities to perceive and manipulate objects. But human beings remain distinctive in their capacity for flexible, creative tool use -- using objects in new ways to act on the world, achieve a goal, or solve a problem. Here we introduce the "Tools" game, a simple but challenging domain for studying this behavior in human and artificial agents. Players place objects in a dynamic scene to accomplish a goal that can only be achieved if those objects interact with other scene elements in appropriate ways: for instance, launching, blocking, supporting or tipping them. Only a few attempts are permitted, requiring rapid trial-and-error learning if a solution is not found at first. We propose a "Sample, Simulate, Update" (SSUP) framework for modeling how people solve these challenges, based on exploiting rich world knowledge to sample actions that would lead to successful outcomes, simulate candidate actions before trying them out, and update beliefs about which tools and actions are best in a rapid learning loop. SSUP captures human performance well across 20 levels of the Tools game, and fits significantly better than alternate accounts based on deep reinforcement learning or learning the simulator parameters online. We discuss how the Tools challenge might guide the development of better physical reasoning agents in AI, as well as better accounts of human physical reasoning and tool use.

Deep reinforcement learning has achieved great successes in recent years, but there are still open challenges, such as convergence to locally optimal policies and sample inefficiency. In this paper, we contribute a novel self-supervised auxiliary task, i.e., Terminal Prediction (TP), estimating temporal closeness to terminal states for episodic tasks. The intuition is to help representation learning by letting the agent predict how close it is to a terminal state, while learning its control policy. Although TP could be integrated with multiple algorithms, this paper focuses on Asynchronous Advantage Actor-Critic (A3C) and demonstrating the advantages of A3C-TP. Our extensive evaluation includes: a set of Atari games, the BipedalWalker domain, and a mini version of the recently proposed multi-agent Pommerman game. Our results on Atari games and the BipedalWalker domain suggest that A3C-TP outperforms standard A3C in most of the tested domains and in others it has similar performance. In Pommerman, our proposed method provides significant improvement both in learning efficiency and converging to better policies against different opponents.

This paper tackles the problem of learning a questioner in the goal-oriented visual dialog task. Several previous works adopt model-free reinforcement learning. Most pretrain the model from a finite set of human-generated data. We argue that using limited demonstrations to kick-start the questioner is insufficient due to the large policy search space. Inspired by a recently proposed information theoretic approach, we develop two analytic experts to serve as a source of high-quality demonstrations for imitation learning. We then take advantage of reinforcement learning to refine the model towards the goal-oriented objective. Experimental results on the GuessWhat?! dataset show that our method has the combined merits of imitation and reinforcement learning, achieving the state-of-the-art performance.

Decision support systems (e.g., for ecological conservation) and autonomous systems (e.g., adaptive controllers in smart cities) start to be deployed in real applications. Although their operations often impact many users or stakeholders, no fairness consideration is generally taken into account in their design, which could lead to completely unfair outcomes for some users or stakeholders. To tackle this issue, we advocate for the use of social welfare functions that encode fairness and present this general novel problem in the context of (deep) reinforcement learning, although it could possibly be extended to other machine learning tasks.

This paper proposes a method for learning a trajectory-conditioned policy to imitate diverse demonstrations from the agent's own past experiences. We demonstrate that such self-imitation drives exploration in diverse directions and increases the chance of finding a globally optimal solution in reinforcement learning problems, especially when the reward is sparse and deceptive. Our method significantly outperforms existing self-imitation learning and count-based exploration methods on various sparse-reward reinforcement learning tasks with local optima. In particular, we report a state-of-the-art score of more than 25,000 points on Montezuma's Revenge without using expert demonstrations or resetting to arbitrary states.