Human gaze is known to be an intention-revealing signal in human demonstrations of tasks. In this work, we use gaze cues from human demonstrators to enhance the performance of state-of-the-art inverse reinforcement learning (IRL) and behavior cloning (BC) algorithms. We propose a novel approach for utilizing gaze data in a computationally efficient manner --- encoding the human's attention as part of an auxiliary loss function, without adding any additional learnable parameters to those models and without requiring gaze data at test time. The auxiliary loss encourages a network to have convolutional activations in regions where the human's gaze fixated. We show how to augment any existing convolutional architecture with our auxiliary gaze loss (coverage-based gaze loss or CGL) that can guide learning toward a better reward function or policy. We show that our proposed approach consistently improves performance of both BC and IRL methods on a variety of Atari games. We also compare against two baseline methods for utilizing gaze data with imitation learning methods. Our approach outperforms a baseline method, called gaze-modulated dropout (GMD), and is comparable to another method (AGIL) which uses gaze as input to the network and thus increases the amount of learnable parameters.

Model-free deep reinforcement learning algorithms are troubled with poor sample efficiency -- learning reliable policies generally requires a vast amount of interaction with the environment. One hypothesis is that catastrophic interference between various segments within the environment is an issue. In this paper, we perform a large-scale empirical study on the presence of catastrophic interference in the Arcade Learning Environment and find that learning particular game segments frequently degrades performance on previously learned segments. In what we term the Memento observation, we show that an identically parameterized agent spawned from a state where the original agent plateaued, reliably makes further progress. This phenomenon is general -- we find consistent performance boosts across architectures, learning algorithms and environments. Our results indicate that eliminating catastrophic interference can contribute towards improved performance and data efficiency of deep reinforcement learning algorithms.

We study the estimation of risk-sensitive policies in reinforcement learning problems defined by a Markov Decision Process (MDPs) whose state and action spaces are countably finite. Prior efforts are predominately afflicted by computational challenges associated with the fact that risk-sensitive MDPs are time-inconsistent. To ameliorate this issue, we propose a new definition of risk, which we call caution, as a penalty function added to the dual objective of the linear programming (LP) formulation of reinforcement learning. The caution measures the distributional risk of a policy, which is a function of the policy's long-term state occupancy distribution. To solve this problem in an online model-free manner, we propose a stochastic variant of primal-dual method that uses Kullback-Lieber (KL) divergence as its proximal term. We establish that the number of iterations/samples required to attain approximately optimal solutions of this scheme matches tight dependencies on the cardinality of the state and action spaces, but differs in its dependence on the infinity norm of the gradient of the risk measure. Experiments demonstrate the merits of this approach for improving the reliability of reward accumulation without additional computational burdens.

Delusional bias is a fundamental source of error in approximate Q-learning. To date, the only techniques that explicitly address delusion require comprehensive search using tabular value estimates. In this paper, we develop efficient methods to mitigate delusional bias by training Q-approximators with labels that are "consistent" with the underlying greedy policy class. We introduce a simple penalization scheme that encourages Q-labels used across training batches to remain (jointly) consistent with the expressible policy class. We also propose a search framework that allows multiple Q-approximators to be generated and tracked, thus mitigating the effect of premature (implicit) policy commitments. Experimental results demonstrate that these methods can improve the performance of Q-learning in a variety of Atari games, sometimes dramatically.

Many AI problems, in robotics and other domains, are goal-based, essentially seeking trajectories leading to various goal states. Reinforcement learning (RL), building on Bellman's optimality equation, naturally optimizes for a single goal, yet can be made multi-goal by augmenting the state with the goal. Instead, we propose a new RL framework, derived from a dynamic programming equation for the all pairs shortest path (APSP) problem, which naturally solves multi-goal queries. We show that this approach has computational benefits for both standard and approximate dynamic programming. Interestingly, our formulation prescribes a novel protocol for computing a trajectory: instead of predicting the next state given its predecessor, as in standard RL, a goal-conditioned trajectory is constructed by first predicting an intermediate state between start and goal, partitioning the trajectory into two. Then, recursively, predicting intermediate points on each sub-segment, until a complete trajectory is obtained. We call this trajectory structure a sub-goal tree. Building on it, we additionally extend the policy gradient methodology to recursively predict sub-goals, resulting in novel goal-based algorithms. Finally, we apply our method to neural motion planning, where we demonstrate significant improvements compared to standard RL on navigating a 7-DoF robot arm between obstacles.

In visual planning (VP), an agent learns to plan goal-directed behavior from observations of a dynamical system obtained offline, e.g., images obtained from self-supervised robot interaction. Most previous works on VP approached the problem by planning in a learned latent space, resulting in low-quality visual plans, and difficult training algorithms. Here, instead, we propose a simple VP method that plans directly in image space and displays competitive performance. We build on the semi-parametric topological memory (SPTM) method: image samples are treated as nodes in a graph, the graph connectivity is learned from image sequence data, and planning can be performed using conventional graph search methods. We propose two modifications on SPTM. First, we train an energy-based graph connectivity function using contrastive predictive coding that admits stable training. Second, to allow zero-shot planning in new domains, we learn a conditional VAE model that generates images given a context of the domain, and use these hallucinated samples for building the connectivity graph and planning. We show that this simple approach significantly outperform the state-of-the-art VP methods, in terms of both plan interpretability and success rate when using the plan to guide a trajectory-following controller. Interestingly, our method can pick up non-trivial visual properties of objects, such as their geometry, and account for it in the plans.

The ability to handle single molecules as effectively as macroscopic building-blocks would enable the construction of complex supramolecular structures that are not accessible by self-assembly. The fundamental challenges on the way towards this goal are the uncontrolled variability and poor observability of atomic-scale conformations. Here, we present a strategy to work around both obstacles, and demonstrate autonomous robotic nanofabrication by manipulating single molecules. Our approach employs reinforcement learning (RL), which is able to learn solution strategies even in the face of large uncertainty and with sparse feedback. However, to be useful for autonomous nanofabrication, standard RL algorithms need to be adapted to cope with the limited training opportunities available. We demonstrate the potential of our RL approach by applying it to an exemplary task of subtractive manufacturing, the removal of individual molecules from a molecular layer using a scanning probe microscope (SPM). Our RL agent reaches an excellent performance level, enabling us to automate a task which previously had to be performed by a human. We anticipate that our work opens the way towards autonomous agents for the robotic construction of functional supramolecular structures with speed, precision and perseverance beyond our current capabilities.

Properly used, positive reinforcement is extremely powerful. Tic-Tac-Toe is a simple game. If both sides play perfectly, neither can win. But if one plays imperfectly, the other can exploit the flaws in the other's strategy. Does that sound a little like trading? Reinforcement learning is a machine learning paradigm that can learn behavior to achieve maximum reward in complex dynamic environments, as simple as Tic-Tac-Toe, or as complex as Go, and options trading. In this post, we will try to explain what reinforcement learning is, share code to apply it, and references to learn more about it.

We study minimax methods for off-policy evaluation (OPE) using value-functions and marginalized importance weights. Despite that they hold promises of overcoming the exponential variance in traditional importance sampling, several key problems remain: (1) They require function approximation and are generally biased. For the sake of trustworthy OPE, is there anyway to quantify the biases? (2) They are split into two styles ("weight-learning" vs "value-learning"). Can we unify them? In this paper we answer both questions positively. By slightly altering the derivation of previous methods (one from each style; Uehara et al., 2019), we unify them into a single confidence interval (CI) that automatically comes with a special type of double robustness: when either the value-function or importance weight class is well-specified, the CI is valid and its length quantifies the misspecification of the other class. We can also tell which class is misspecified, which provides useful diagnostic information for the design of function approximation. Our CI also provides a unified view of and new insights to some recent methods: for example, one side of the CI recovers a version of AlgaeDICE (Nachum et al., 2019b), and we show that the two sides need to be used together and either alone may incur doubled approximation error as a point estimate. We further examine the potential of applying these bounds to two long-standing problems: off-policy policy optimization with poor data coverage (i.e., exploitation), and systematic exploration. With a well-specified value-function class, we show that optimizing the lower and the upper bounds lead to effective exploitation and exploration, respectively. Our results also suggests an interesting assymetry between exploration and exploitation, that the former might require substantially weaker realizability assumptions than the latter.

For a robotic grasping task in which diverse unseen target objects exist in a cluttered environment, some deep learning-based methods have achieved state-of-the-art results using visual input directly. In contrast, actor-critic deep reinforcement learning (RL) methods typically perform very poorly when grasping diverse objects, especially when learning from raw images and sparse rewards. To make these RL techniques feasible for vision-based grasping tasks, we employ state representation learning (SRL), where we encode essential information first for subsequent use in RL. However, typical representation learning procedures are unsuitable for extracting pertinent information for learning the grasping skill, because the visual inputs for representation learning, where a robot attempts to grasp a target object in clutter, are extremely complex. We found that preprocessing based on the disentanglement of a raw input image is the key to effectively capturing a compact representation. This enables deep RL to learn robotic grasping skills from highly varied and diverse visual inputs. We demonstrate the effectiveness of this approach with varying levels of disentanglement in a realistic simulated environment.