Representation learning in Reinforcement Learning (RL) has gained increasing attention in recent years from both theoretical and empirical research communities (Schwarzer et al., 2020; Laskin et al., 2020) due to its potential in enabling sample-efficient non-linear function approximation, the benefits in multitask settings (Zhang et al., 2020; Yang et al., 2022; Sodhani et al., 2021), and the potential to leverage advances on representation learning in related areas such as computer vision and natural language processing. Despite this interest, there remains a gap between the theoretical and empirical literature, where the theoretically sound methods are seldom evaluated or even implemented and often rely on strong assumptions, while the empirical techniques are not backed with any theoretical guarantees even under stylistic assumptions. This leaves open the key challenge of designing representation learning methods that are both theoretically sound and empirically effective. In this work, we tackle this challenge for a special class of problems called Block MDPs, where the high dimensional and rich observations of the agent are generated from certain latent states and there exists some fixed, but unknown mapping from observations to the latent states (each observation is generated only by one latent state). Prior works (Dann et al., 2018; Du et al., 2019; Misra et al., 2020; Zhang et al., 2020; Sodhani et al., 2021) have motivated the Block MDP model through scenarios such as navigation tasks and image based robotics tasks where the observations can often be reasonably mapped to the latent physical location and states.

To accumulate knowledge and improve its policy of behaviour, a reinforcement learning agent can learn `off-policy' about policies that differ from the policy used to generate its experience. This is important to learn counterfactuals, or because the experience was generated out of its own control. However, off-policy learning is non-trivial, and standard reinforcement-learning algorithms can be unstable and divergent. In this paper we discuss a novel family of off-policy prediction algorithms which are convergent by construction. The idea is to first learn on-policy about the data-generating behaviour, and then bootstrap an off-policy value estimate on this on-policy estimate, thereby constructing a value estimate that is partially off-policy. This process can be repeated to build a chain of value functions, each time bootstrapping a new estimate on the previous estimate in the chain. Each step in the chain is stable and hence the complete algorithm is guaranteed to be stable. Under mild conditions this comes arbitrarily close to the off-policy TD solution when we increase the length of the chain. Hence it can compute the solution even in cases where off-policy TD diverges. We prove that the proposed scheme is convergent and corresponds to an iterative decomposition of the inverse key matrix. Furthermore it can be interpreted as estimating a novel objective -- that we call a `k-step expedition' -- of following the target policy for finitely many steps before continuing indefinitely with the behaviour policy. Empirically we evaluate the idea on challenging MDPs such as Baird's counter example and observe favourable results.

The development of self-propelled particles at the micro-and the nanoscale has sparked a huge potential for future applications in active matter physics, microsurgery, and targeted drug delivery. However, while the latter applications provoke the quest on how to optimally navigate towards a target, such as e.g. a cancer cell, there is still no simple way known to determine the optimal route in sufficiently complex environments. Here we develop a machine learning-based approach that allows us, for the first time, to determine the asymptotically optimal path of a selfpropelled agent which can freely steer in complex environments. Our method hinges on policy gradient-based deep reinforcement learning techniques and, crucially, does not require any reward shaping or heuristics. The presented method provides a powerful alternative to current analytical methods to calculate optimal trajectories and opens a route towards a universal path planner for future intelligent active particles. Keywords: Active Matter Physics, Colloids, Soft Matter Physics, Optimization, Microswimmers, Optimal Navigation, Reinforcement Learning Introduction The problem of finding suitable navigation strategies is of great interest to applications ranging from motion planning for autonomous underwater vehicles, ocean gliders [1-4], and aerial vehicles [5-7] to microorganisms searching for food and prey [8,9] and striving for survival in complex environments [10, 11]. One important class of path planning problems which is currently attracting a rapidly increasing attention is centered around the quest for the optimal trajectory allowing an active particle, which can freely steer but cannot control its speed, to reach a given target in a complex environment. This active particle navigation (APN) problem is relevant both for biological swimmers like fish or for turtles on the way to their breeding grounds [12,13] and for future applications of synthetic microswimmers [14] such as targeted drug [15-17] and gene delivery [18,19] or microsurgery [20].

Differentiable simulators promise faster computation time for reinforcement learning by replacing zeroth-order gradient estimates of a stochastic objective with an estimate based on first-order gradients. However, it is yet unclear what factors decide the performance of the two estimators on complex landscapes that involve long-horizon planning and control on physical systems, despite the crucial relevance of this question for the utility of differentiable simulators. We show that characteristics of certain physical systems, such as stiffness or discontinuities, may compromise the efficacy of the first-order estimator, and analyze this phenomenon through the lens of bias and variance. We additionally propose an $\alpha$-order gradient estimator, with $\alpha \in [0,1]$, which correctly utilizes exact gradients to combine the efficiency of first-order estimates with the robustness of zero-order methods. We demonstrate the pitfalls of traditional estimators and the advantages of the $\alpha$-order estimator on some numerical examples.

Much of recent Deep Reinforcement Learning success is owed to the neural architecture's potential to learn and use effective internal representations of the world. While many current algorithms access a simulator to train with a large amount of data, in realistic settings, including while playing games that may be played against people, collecting experience can be quite costly. In this paper, we introduce a deep reinforcement learning architecture whose purpose is to increase sample efficiency without sacrificing performance. We design this architecture by incorporating advances achieved in recent years in the field of Natural Language Processing and Computer Vision. Specifically, we propose a visually attentive model that uses transformers to learn a self-attention mechanism on the feature maps of the state representation, while simultaneously optimizing return. We demonstrate empirically that this architecture improves sample complexity for several Atari environments, while also achieving better performance in some of the games.

Optimization is a ubiquitous modeling tool and is often deployed in settings which repeatedly solve similar instances of the same problem. Amortized optimization methods use learning to predict the solutions to problems in these settings. This leverages the shared structure between similar problem instances. In this tutorial, we will discuss the key design choices behind amortized optimization, roughly categorizing 1) models into fully-amortized and semi-amortized approaches, and 2) learning methods into regression-based and objectivebased. We then view existing applications through these foundations to draw connections between them, including for manifold optimization, variational inference, sparse coding, meta-learning, control, reinforcement learning, convex optimization, and deep equilibrium networks. This framing enables us easily see, for example, that the amortized inference in variational autoencoders is conceptually identical to value gradients in control and reinforcement learning as they both use fully-amortized models with an objective-based loss.

Machine learning has the potential to automate molecular design and drastically accelerate the discovery of new functional compounds. Towards this goal, generative models and reinforcement learning (RL) using string and graph representations have been successfully used to search for novel molecules. However, these approaches are limited since their representations ignore the three-dimensional (3D) structure of molecules. In fact, geometry plays an important role in many applications in inverse molecular design, especially in drug discovery. Thus, it is important to build models that can generate molecular structures in 3D space based on property-oriented geometric constraints. To address this, one approach is to generate molecules as 3D point clouds by sequentially placing atoms at locations in space -- this allows the process to be guided by physical quantities such as energy or other properties. However, this approach is inefficient as placing individual atoms makes the exploration unnecessarily deep, limiting the complexity of molecules that can be generated. Moreover, when optimizing a molecule, organic and medicinal chemists use known fragments and functional groups, not single atoms. We introduce a novel RL framework for scalable 3D design that uses a hierarchical agent to build molecules by placing molecular substructures sequentially in 3D space, thus attempting to build on the existing human knowledge in the field of molecular design. In a variety of experiments with different substructures, we show that our agent, guided only by energy considerations, can efficiently learn to produce molecules with over 100 atoms from many distributions including drug-like molecules, organic LED molecules, and biomolecules.

The imitation learning research community has recently made significant progress towards the goal of enabling artificial agents to imitate behaviors from video demonstrations alone. However, current state-of-the-art approaches developed for this problem exhibit high sample complexity due, in part, to the high-dimensional nature of video observations. Towards addressing this issue, we introduce here a new algorithm called Visual Generative Adversarial Imitation from Observation using a State Observer VGAIfO-SO. At its core, VGAIfO-SO seeks to address sample inefficiency using a novel, self-supervised state observer, which provides estimates of lower-dimensional proprioceptive state representations from high-dimensional images. We show experimentally in several continuous control environments that VGAIfO-SO is more sample efficient than other IfO algorithms at learning from video-only demonstrations and can sometimes even achieve performance close to the Generative Adversarial Imitation from Observation (GAIfO) algorithm that has privileged access to the demonstrator's proprioceptive state information.

Contextual bandits are commonly used for sequential decision-making with finitely many control actions. In this setting, available context observations can be utilized in a tractable way, thanks to the linearity of the relationship between the reward and the context vectors. The arms provide rewards depending on the contexts that represent their individual characteristics. The range of real-world applications is notably extensive, including personalized recommendations for Mobile Context-Aware Recommender Systems and mobile-health interventions [1, 2, 3]. To get satisfactory performances in bandits, the exploration-exploitation trade-off must be addressed. The theoretical analysis of efficient policies for the multi-armed bandits goes back to algorithms that decide based on Upper-Confident-Bounds (UCB) [4]. In fact, UCB employs an optimistic approximate of the unknown reward based on the history of observations, to allow an appropriate degree of exploration. Further theoretical results for UCB in contextual bandits, as well as in other settings, are available in the literature [5, 6, 7, 8, 9]. Posterior sampling is another ubiquitous reinforcement learning algorithm that effectively balances exploitation versus exploration.

Bayesian approaches developed to solve the optimal design of sequential experiments are mathematically elegant but computationally challenging. Recently, techniques using amortization have been proposed to make these Bayesian approaches practical, by training a parameterized policy that proposes designs efficiently at deployment time. However, these methods may not sufficiently explore the design space, require access to a differentiable probabilistic model and can only optimize over continuous design spaces. Here, we address these limitations by showing that the problem of optimizing policies can be reduced to solving a Markov decision process (MDP). We solve the equivalent MDP with modern deep reinforcement learning techniques. Our experiments show that our approach is also computationally efficient at deployment time and exhibits state-of-the-art performance on both continuous and discrete design spaces, even when the probabilistic model is a black box.