We study active preference learning as a framework for intuitively specifying the behaviour of autonomous robots. In active preference learning, a user chooses the preferred behaviour from a set of alternatives, from which the robot learns the user's preferences, modeled as a parameterized cost function. Previous approaches present users with alternatives that minimize the uncertainty over the parameters of the cost function. However, different parameters might lead to the same optimal behaviour; as a consequence the solution space is more structured than the parameter space. We exploit this by proposing a query selection that greedily reduces the maximum error ratio over the solution space. In simulations we demonstrate that the proposed approach outperforms other state of the art techniques in both learning efficiency and ease of queries for the user. Finally, we show that evaluating the learning based on the similarities of solutions instead of the similarities of weights allows for better predictions for different scenarios.

The overestimation bias is one of the major impediments to accurate off-policy learning. This paper investigates a novel way to alleviate the overestimation bias in a continuous control setting. Our method---Truncated Quantile Critics, TQC,---blends three ideas: distributional representation of a critic, truncation of critics prediction, and ensembling of multiple critics. Distributional representation and truncation allow for arbitrary granular overestimation control, while ensembling provides additional score improvements. TQC outperforms the current state of the art on all environments from the continuous control benchmark suite, demonstrating 25% improvement on the most challenging Humanoid environment.

In this paper we propose Policy Synthesis under probabilistic Constraints (PSyCo), a systematic engineering method for synthesizing safe policies under probabilistic constraints with reinforcement learning and Bayesian model checking. As an implementation of PSyCo we introduce Safe Neural Evolutionary Strategies (SNES). SNES leverages Bayesian model checking while learning to adjust the Lagrangian of a constrained optimization problem derived from a PSyCo specification. We empirically evaluate SNES' ability to synthesize feasible policies in settings with formal safety requirements.

MineRL 2019 competition challenged participants to train sample-efficient agents to play Minecraft, by using a dataset of human gameplay and a limit number of steps the environment. We approached this task with behavioural cloning by predicting what actions human players would take, and reached fifth place in the final ranking. Despite being a simple algorithm, we observed the performance of such an approach can vary significantly, based on when the training is stopped. In this paper, we detail our submission to the competition, run further experiments to study how performance varied over training and study how different engineering decisions affected these results.

As an important type of reinforcement learning algorithms, actor-critic (AC) and natural actor-critic (NAC) algorithms are often executed in two ways for finding optimal policies. In the first nested-loop design, actor's one update of policy is followed by an entire loop of critic's updates of the value function, and the finite-sample analysis of such AC and NAC algorithms have been recently well established. The second two time-scale design, in which actor and critic update simultaneously but with different learning rates, has much fewer tuning parameters than the nested-loop design and is hence substantially easier to implement. Although two time-scale AC and NAC have been shown to converge in the literature, the finite-sample convergence rate has not been established. In this paper, we provide the first such non-asymptotic convergence rate for two time-scale AC and NAC under Markovian sampling and with actor having general policy class approximation. We show that two time-scale AC requires the overall sample complexity at the order of $\mathcal{O}(\epsilon^{-2.5}\log^3(\epsilon^{-1}))$ to attain an $\epsilon$-accurate stationary point, and two time-scale NAC requires the overall sample complexity at the order of $\mathcal{O}(\epsilon^{-4}\log^2(\epsilon^{-1}))$ to attain an $\epsilon$-accurate global optimal point. We develop novel techniques for bounding the bias error of the actor due to dynamically changing Markovian sampling and for analyzing the convergence rate of the linear critic with dynamically changing base functions and transition kernel.

Reinforcement learning methods have been used to compute dialog policies from language-based interaction experiences. Efficiency is of particular importance in dialog policy learning, because of the considerable cost of interacting with people, and the very poor user experience from low-quality conversations. Aiming at improving the efficiency of dialog policy learning, we develop algorithm LHUA (Learning with Hindsight, User modeling, and Adaptation) that, for the first time, enables dialog agents to adaptively learn with hindsight from both simulated and real users. Simulation and hindsight provide the dialog agent with more experience and more (positive) reinforcements respectively. Experimental results suggest that, in success rate and policy quality, LHUA outperforms competitive baselines from the literature, including its no-simulation, no-adaptation, and no-hindsight counterparts.

We study episodic reinforcement learning in Markov decision processes when the agent receives additional feedback per step in the form of several transition observations. Such additional observations are available in a range of tasks through extended sensors or prior knowledge about the environment (e.g., when certain actions yield similar outcome). We formalize this setting using a feedback graph over state-action pairs and show that model-based algorithms can leverage the additional feedback for more sample-efficient learning. We give a regret bound that, ignoring logarithmic factors and lower-order terms, depends only on the size of the maximum acyclic subgraph of the feedback graph, in contrast with a polynomial dependency on the number of states and actions in the absence of a feedback graph. Finally, we highlight challenges when leveraging a small dominating set of the feedback graph as compared to the bandit setting and propose a new algorithm that can use knowledge of such a dominating set for more sample-efficient learning of a near-optimal policy.

In this paper we introduce plan2vec, an unsupervised representation learning approach that is inspired by reinforcement learning. Plan2vec constructs a weighted graph on an image dataset using near-neighbor distances, and then extrapolates this local metric to a global embedding by distilling path-integral over planned path. When applied to control, plan2vec offers a way to learn goal-conditioned value estimates that are accurate over long horizons that is both compute and sample efficient. We demonstrate the effectiveness of plan2vec on one simulated and two challenging real-world image datasets. Experimental results show that plan2vec successfully amortizes the planning cost, enabling reactive planning that is linear in memory and computation complexity rather than exhaustive over the entire state space.

It was not long ago that the world watched World Chess Champion Garry Kasparov lose a decisive match against a supercomputer. IBM's Deep Blue embodied the state of the art in the late 1990s, when a machine defeating a world (human) champion at a complex game such as chess was still unheard of. Fast-forward to today, and not only have supercomputers greatly surpassed Deep Blue in chess, they have managed to achieve superhuman performance in a string of other games, often much more complex than chess, ranging from Go to Dota to classic Atari titles. Many of these games have been mastered just in the last five years, pointing to a pace of innovation much quicker than the two decades prior. Recently, Google released work on Agent57, which for the first time showcased superior performance over existing benchmarks across all 57 Atari 2600 games. The class of AI algorithms underlying these feats -- deep-reinforcement learning -- has demonstrated the ability to learn at very high levels in constrained domains, such as the ones offered by games.

Reinforcement learning has seen a great deal of success in solving complex decision making problems ranging from robotics to games to supply chain management to recommender systems. Despite their success, deep reinforcement learning algorithms can be exceptionally difficult to use, due to unstable training, sensitivity to hyperparameters, and generally unpredictable and poorly understood convergence properties. Multiple explanations, and corresponding solutions, have been proposed for improving the stability of such methods, and we have seen good progress over the last few years on these algorithms. In this blog post, we will dive deep into analyzing a central and underexplored reason behind some of the problems with the class of deep RL algorithms based on dynamic programming, which encompass the popular DQN and soft actor-critic (SAC) algorithms – the detrimental connection between data distributions and learned models. Before diving deep into a description of this problem, let us quickly recap some of the main concepts in dynamic programming.