Optimal scaling has been well studied for Metropolis-Hastings (M-H) algorithms in continuous spaces, but a similar understanding has been lacking in discrete spaces. Recently, a family of locally balanced proposals (LBP) for discrete spaces has been proved to be asymptotically optimal, but the question of optimal scaling has remained open. In this paper, we establish, for the first time, that the efficiency of M-H in discrete spaces can also be characterized by an asymptotic acceptance rate that is independent of the target distribution. Moreover, we verify, both theoretically and empirically, that the optimal acceptance rates for LBP and random walk Metropolis (RWM) are $0.574$ and $0.234$ respectively. These results also help establish that LBP is asymptotically $O(N^\frac{2}{3})$ more efficient than RWM with respect to model dimension $N$. Knowledge of the optimal acceptance rate allows one to automatically tune the neighborhood size of a proposal distribution in a discrete space, directly analogous to step-size control in continuous spaces. We demonstrate empirically that such adaptive M-H sampling can robustly improve sampling in a variety of target distributions in discrete spaces, including training deep energy based models.

Interactive reinforcement learning can effectively facilitate the agent training via human feedback. However, such methods often require the human teacher to know what is the correct action that the agent should take. In other words, if the human teacher is not always reliable, then it will not be consistently able to guide the agent through its training. In this paper, we propose a more effective interactive reinforcement learning system by introducing multiple trainers, namely Multi-Trainer Interactive Reinforcement Learning (MTIRL), which could aggregate the binary feedback from multiple non-perfect trainers into a more reliable reward for an agent training in a reward-sparse environment. In particular, our trainer feedback aggregation experiments show that our aggregation method has the best accuracy when compared with the majority voting, the weighted voting, and the Bayesian method. Finally, we conduct a grid-world experiment to show that the policy trained by the MTIRL with the review model is closer to the optimal policy than that without a review model.

Recent years have seen embodied visual navigation advance in two distinct directions: (i) in equipping the AI agent to follow natural language instructions, and (ii) in making the navigable world multimodal, e.g., audio-visual navigation. However, the real world is not only multimodal, but also often complex, and thus in spite of these advances, agents still need to understand the uncertainty in their actions and seek instructions to navigate. To this end, we present AVLEN~ -- an interactive agent for Audio-Visual-Language Embodied Navigation. Similar to audio-visual navigation tasks, the goal of our embodied agent is to localize an audio event via navigating the 3D visual world; however, the agent may also seek help from a human (oracle), where the assistance is provided in free-form natural language. To realize these abilities, AVLEN uses a multimodal hierarchical reinforcement learning backbone that learns: (a) high-level policies to choose either audio-cues for navigation or to query the oracle, and (b) lower-level policies to select navigation actions based on its audio-visual and language inputs. The policies are trained via rewarding for the success on the navigation task while minimizing the number of queries to the oracle. To empirically evaluate AVLEN, we present experiments on the SoundSpaces framework for semantic audio-visual navigation tasks. Our results show that equipping the agent to ask for help leads to a clear improvement in performance, especially in challenging cases, e.g., when the sound is unheard during training or in the presence of distractor sounds.

A/B tests) are a powerful tool for estimating treatment effects, to inform decisions making in business, healthcare and other applications. In many problems, the treatment has a lasting effect that evolves over time. A limitation with randomized experiments is that they do not easily extend to measure long-term effects, since running long experiments is time-consuming and expensive. In this paper, we take a reinforcement learning (RL) approach that estimates the average reward in a Markov process. Motivated by real-world scenarios where the observed state transition is nonstationary, we develop a new algorithm for a class of nonstationary problems, and demonstrate promising results in two synthetic datasets and one online store dataset.

Algorithms for text-generation in dialogue can be misguided. For example, in task-oriented settings, reinforcement learning that optimizes only task-success can lead to abysmal lexical diversity. We hypothesize this is due to poor theoretical understanding of the objectives in text-generation and their relation to the learning process (i.e., model training). To this end, we propose a new theoretical framework for learning to generate text in dialogue. Compared to existing theories of learning, our framework allows for analysis of the multi-faceted goals inherent to text-generation. We use our framework to develop theoretical guarantees for learners that adapt to unseen data. As an example, we apply our theory to study data-shift within a cooperative learning algorithm proposed for the GuessWhat?! visual dialogue game. From this insight, we propose a new algorithm, and empirically, we demonstrate our proposal improves both task-success and human-likeness of the generated text. Finally, we show statistics from our theory are empirically predictive of multiple qualities of the generated dialogue, suggesting our theory is useful for model-selection when human evaluations are not available.

We propose a family of First Hitting Diffusion Models (FHDM), deep generative models that generate data with a diffusion process that terminates at a random first hitting time. This yields an extension of the standard fixed-time diffusion models that terminate at a pre-specified deterministic time. Although standard diffusion models are designed for continuous unconstrained data, FHDM is naturally designed to learn distributions on continuous as well as a range of discrete and structure domains. Moreover, FHDM enables instance-dependent terminate time and accelerates the diffusion process to sample higher quality data with fewer diffusion steps. Technically, we train FHDM by maximum likelihood estimation on diffusion trajectories augmented from observed data with conditional first hitting processes (i.e., bridge) derived based on Doob's $h$-transform, deviating from the commonly used time-reversal mechanism. We apply FHDM to generate data in various domains such as point cloud (general continuous distribution), climate and geographical events on earth (continuous distribution on the sphere), unweighted graphs (distribution of binary matrices), and segmentation maps of 2D images (high-dimensional categorical distribution). We observe considerable improvement compared with the state-of-the-art approaches in both quality and speed.

Getting robots to navigate to multiple objects autonomously is essential yet difficult in robot applications. One of the key challenges is how to explore environments efficiently with camera sensors only. Existing navigation methods mainly focus on fixed cameras and few attempts have been made to navigate with active cameras. As a result, the agent may take a very long time to perceive the environment due to limited camera scope. In contrast, humans typically gain a larger field of view by looking around for a better perception of the environment. How to make robots perceive the environment as efficiently as humans is a fundamental problem in robotics. In this paper, we consider navigating to multiple objects more efficiently with active cameras. Specifically, we cast moving camera to a Markov Decision Process and reformulate the active camera problem as a reinforcement learning problem. However, we have to address two new challenges: 1) how to learn a good camera policy in complex environments and 2) how to coordinate it with the navigation policy. To address these, we carefully design a reward function to encourage the agent to explore more areas by moving camera actively. Moreover, we exploit human experience to infer a rule-based camera action to guide the learning process. Last, to better coordinate two kinds of policies, the camera policy takes navigation actions into account when making camera moving decisions. Experimental results show our camera policy consistently improves the performance of multi-object navigation over four baselines on two datasets.

Accurate localization is a critical requirement for most robotic tasks. The main body of existing work is focused on passive localization in which the motions of the robot are assumed given, abstracting from their influence on sampling informative observations. While recent work has shown the benefits of learning motions to disambiguate the robot's poses, these methods are restricted to granular discrete actions and directly depend on the size of the global map. We propose Active Particle Filter Networks (APFN), an approach that only relies on local information for both the likelihood evaluation as well as the decision making. To do so, we couple differentiable particle filters with a reinforcement learning agent that attends to the most relevant parts of the map. The resulting approach inherits the computational benefits of particle filters and can directly act in continuous action spaces while remaining fully differentiable and thereby end-to-end optimizable as well as agnostic to the input modality. We demonstrate the benefits of our approach with extensive experiments in photorealistic indoor environments built from real-world 3D scanned apartments. Videos and code are available at http://apfn.cs.uni-freiburg.de.

The discovery of novel inhibitor molecules for emerging drug-target proteins is widely acknowledged as a challenging inverse design problem: Exhaustive exploration of the vast chemical search space is impractical, especially when the target structure or active molecules are unknown. Here we validate experimentally the broad utility of a deep generative framework trained at-scale on protein sequences, small molecules, and their mutual interactions -- that is unbiased toward any specific target. As demonstrators, we consider two dissimilar and relevant SARS-CoV-2 targets: the main protease and the spike protein (receptor binding domain, RBD). To perform target-aware design of novel inhibitor molecules, a protein sequence-conditioned sampling on the generative foundation model is performed. Despite using only the target sequence information, and without performing any target-specific adaptation of the generative model, micromolar-level inhibition was observed in in vitro experiments for two candidates out of only four synthesized for each target. The most potent spike RBD inhibitor also exhibited activity against several variants in live virus neutralization assays. These results therefore establish that a single, broadly deployable generative foundation model for accelerated hit discovery is effective and efficient, even in the most general case where neither target structure nor binder information is available.

The uses of robots are changing from static environments in factories to encompass novel concepts such as Human-Robot Collaboration in unstructured settings. Pre-programming all the functionalities for robots becomes impractical, and hence, robots need to learn how to react to new events autonomously, just like humans. However, humans, unlike machines, are naturally skilled in responding to unexpected circumstances based on either experiences or observations. Hence, embedding such anthropoid behaviours into robots entails the development of neuro-cognitive models that emulate motor skills under a robot learning paradigm. Effective encoding of these skills is bound to the proper choice of tools and techniques. This paper studies different motion and behaviour learning methods ranging from Movement Primitives (MP) to Experience Abstraction (EA), applied to different robotic tasks. These methods are scrutinized and then experimentally benchmarked by reconstructing a standard pick-n-place task. Apart from providing a standard guideline for the selection of strategies and algorithms, this paper aims to draw a perspectives on their possible extensions and improvements