While much work in data science to date has focused on algorithmic scale and sophistication, safety -- that is, safeguards against harm -- is a domain no less worth pursuing. This is particularly true in applications like self-driving vehicles, where a machine learning system's poor judgement might contribute to an accident. That's why firms like Intel's Mobileye and Nvidia have proposed frameworks to guarantee safe and logical decision-making, and it's why OpenAI -- the San Francisco-based research firm cofounded by CTO Greg Brockman, chief scientist Ilya Sutskever, and others -- today released Safety Gym. OpenAI describes it as a suite of tools for developing AI that respects safety constraints while training, and for comparing the "safety" of algorithms and the extent to which those algorithms avoid mistakes while learning. Safety Gym is designed for reinforcement learning agents, or AI that's progressively spurred toward goals via rewards (or punishments).

We integrate information-theoretic concepts into the design and analysis of optimistic algorithms and Thompson sampling. By making a connection between information-theoretic quantities and confidence bounds, we obtain results that relate the per-period performance of the agent with its information gain about the environment, thus explicitly characterizing the exploration-exploitation tradeoff. The resulting cumulative regret bound depends on the agent's uncertainty over the environment and quantifies the value of prior information. We show applicability of this approach to several environments, including linear bandits, tabular MDPs, and factored MDPs. These examples demonstrate the potential of a general information-theoretic approach for the design and analysis of reinforcement learning algorithms.

We present new algorithms for computing and approximating bisimulation metrics in Markov Decision Processes (MDPs). Bisimulation metrics are an elegant formalism that capture behavioral equivalence between states and provide strong theoretical guarantees on differences in optimal behaviour. Unfortunately, their computation is expensive and requires a tabular representation of the states, which has thus far rendered them impractical for large problems. In this paper we present a new version of the metric that is tied to a behavior policy in an MDP, along with an analysis of its theoretical properties. We then present two new algorithms for approximating bisimulation metrics in large, deterministic MDPs. The first does so via sampling and is guaranteed to converge to the true metric. The second is a differentiable loss which allows us to learn an approximation even for continuous state MDPs, which prior to this work had not been possible.

Deep networks have enabled reinforcement learning to scale to more complex and challenging domains, but these methods typically require large quantities of training data. An alternative is to use sample-efficient episodic control methods: neuro-inspired algorithms which use non-/semi-parametric models that predict values based on storing and retrieving previously experienced transitions. One way to further improve the sample efficiency of these approaches is to use more principled exploration strategies. In this work, we therefore propose maximum entropy mellowmax episodic control (MEMEC), which samples actions according to a Boltzmann policy with a state-dependent temperature. We demonstrate that MEMEC outperforms other uncertainty- and softmax-based exploration methods on classic reinforcement learning environments and Atari games, achieving both more rapid learning and higher final rewards.

Recently, neuro-inspired episodic control (EC) methods have been developed to overcome the data-inefficiency of standard deep reinforcement learning approaches. Using non-/semi-parametric models to estimate the value function, they learn rapidly, retrieving cached values from similar past states. In realistic scenarios, with limited resources and noisy data, maintaining meaningful representations in memory is essential to speed up the learning and avoid catastrophic forgetting. Unfortunately, EC methods have a large space and time complexity. We investigate different solutions to these problems based on prioritising and ranking stored states, as well as online clustering techniques. We also propose a new dynamic online k-means algorithm that is both computationally-efficient and yields significantly better performance at smaller memory sizes; we validate this approach on classic reinforcement learning environments and Atari games.

A BSTRACT Consider an imitation learning problem that the imitator and the expert have different dynamics models. Most of the current imitation learning methods fail because they focus on imitating actions. We propose a novel state alignment based imitation learning method to train the imitator to follow the state sequences in expert demonstrations as much as possible. The state alignment comes from both local and global perspectives and we combine them into a reinforcement learning framework by a regularized policy update objective. We show the superiority of our method on standard imitation learning settings and imitation learning settings where the expert and imitator have different dynamics models. 1 I NTRODUCTION Learning from demonstrations (imitation learning, abbr. Imitation learning methods can be generally divided into two categories: behavior cloning (BC) and inverse reinforcement learning (IRL). Behavior cloning (Ross et al., 2011b) formulates a supervised learning problem to learn a policy that maps states to actions using demonstration trajectories. Inverse reinforcement learning (Russell, 1998; Ng et al., 2000) tries to find a proper reward function that can induce the given demonstration trajectories. GAIL (Ho & Ermon, 2016) and its variants (Fu et al.; Qureshi et al., 2018; Xiao et al., 2019) are the recently proposed IRL-based methods, which uses a GAN-based reward to align the distribution of state-action pairs between the expert and the imitator.

Making the right decisions when some of the state variables are hidden, involves reasoning about all the possible states of the environment. An agent receiving only partial observations needs to infer the true values of these hidden variables based on the history of experiences. Recent deep reinforcement learning methods use recurrent models to keep track of past information. However, these models are sometimes expensive to train and have convergence difficulties, especially when dealing with high dimensional input spaces. Taking inspiration from influence-based abstraction, we show that effective policies can be learned in the presence of uncertainty by only memorizing a small subset of input variables. We also incorporate a mechanism in our network that learns to automatically choose the important pieces of information that need to be remembered. The results indicate that, by forcing the agent's internal memory to focus on the selected regions while treating the rest of the observable variables as Markovian, we can outperform ordinary recurrent architectures in situations where the amount of information that the agent needs to retain represents a small fraction of the entire observation input. The method also reduces training time and obtains better scores than methods that stack multiple observations to remove partial observability in domains where long-term memory is required.

Reinforcement Learning in domains with sparse rewards is a difficult problem, and a large part of the training process is often spent searching the state space in a more or less random fashion for any learning signals. For control problems, we often have some controller readily available which might be suboptimal but nevertheless solves the problem to some degree. This controller can be used to guide the initial exploration phase of the learning controller towards reward yielding states, reducing the time before refinement of a viable policy can be initiated. In our work, the agent is guided through an auxiliary behaviour cloning loss which is made conditional on a Q-filter, i.e. it is only applied in situations where the critic deems the guiding controller to be better than the agent. The Q-filter provides a natural way to adjust the guidance throughout the training process, allowing the agent to exceed the guiding controller in a manner that is adaptive to the task at hand and the proficiency of the guiding controller. The contribution of this paper lies in identifying shortcomings in previously proposed implementations of the Q-filter concept, and in suggesting some ways these issues can be mitigated. These modifications are tested on the OpenAI Gym Fetch environments, showing clear improvements in adaptivity and yielding increased performance in all robotic environments tested.

We focus on a class of real-world domains, where gathering hierarchical knowledge is required to accomplish a task. Many problems can be represented in this manner, such as network penetration testing, targeted advertising or medical diagnosis. In our formalization, the task is to sequentially request pieces of information about a sample to build the knowledge hierarchy and terminate when suitable. Any of the learned pieces of information can be further analyzed, resulting in a complex and variable action space. We present a combination of techniques in which the knowledge hierarchy is explicitly represented and given to a deep reinforcement learning algorithm as its input. To process the hierarchical input, we employ Hierarchical Multiple-Instance Learning and to cope with the complex action space, we factor it with hierarchical softmax. Our end-to-end differentiable model is trained with A2C, a standard deep reinforcement learning algorithm. We demonstrate the method in a set of seven classification domains, where the task is to achieve the best accuracy with a set budget on the amount of information retrieved. Compared to baseline algorithms, our method achieves not only better results, but also better generalization.

T om Zahavy, Alon Cohen, Haim Kaplan and Yishay Mansour Google Research, Tel Aviv Abstract We consider the applications of the Frank-Wolfe (FW) algorithm for Apprenticeship Learning (AL). In this setting, we are given a Markov Decision Process (MDP) without an explicit reward function. Instead, we observe an expert that acts according to some policy, and the goal is to find a policy whose feature expectations are closest to those of the expert policy. We formulate this problem as finding the projection of the feature expectations of the expert on the feature expectations polytope - the convex hull of the feature expectations of all the deterministic policies in the MDP . We show that this formulation is equivalent to the AL objective and that solving this problem using the FW algorithm is equivalent well-known Projection method of Abbeel and Ng (2004). This insight allows us to analyze AL with tools from convex optimization literature and derive tighter convergence bounds on AL. Specifically, we show that a variation of the FW method that is based on taking "away steps" achieves a linear rate of convergence when applied to AL and that a stochastic version of the FW algorithm can be used to avoid precise estimation of feature expectations. We also experimentally show that this version outperforms the FW baseline. To the best of our knowledge, this is the first work that shows linear convergence rates for AL. 1 Introduction We consider sequential decision making in the Markov decision process (MDP) formalism. Given an MDP, the optimal policy and its value function are characterized by the Bellman equations and can be computed via value or policy iteration. This makes the MDP model useful in problems where we can specify the MDP model (states, actions, reward, transitions) appropriately. However, in many real-world problems, it is often hard to define a reward function, such that the optimal policy with respect to this reward produces the desired behavior.