We study the multi-step off-policy learning approach to distributional RL. Despite the apparent similarity between value-based RL and distributional RL, our study reveals intriguing and fundamental differences between the two cases in the multi-step setting. We identify a novel notion of path-dependent distributional TD error, which is indispensable for principled multi-step distributional RL. The distinction from the value-based case bears important implications on concepts such as backward-view algorithms. Our work provides the first theoretical guarantees on multi-step off-policy distributional RL algorithms, including results that apply to the small number of existing approaches to multi-step distributional RL. In addition, we derive a novel algorithm, Quantile Regression-Retrace, which leads to a deep RL agent QR-DQN-Retrace that shows empirical improvements over QR-DQN on the Atari-57 benchmark. Collectively, we shed light on how unique challenges in multi-step distributional RL can be addressed both in theory and practice.

Experience replay is an essential component in deep reinforcement learning (DRL), which stores the experiences and generates experiences for the agent to learn in real time. Recently, prioritized experience replay (PER) has been proven to be powerful and widely deployed in DRL agents. However, implementing PER on traditional CPU or GPU architectures incurs significant latency overhead due to its frequent and irregular memory accesses. This paper proposes a hardware-software co-design approach to design an associative memory (AM) based PER, AMPER, with an AM-friendly priority sampling operation. AMPER replaces the widely-used time-costly tree-traversal-based priority sampling in PER while preserving the learning performance. Further, we design an in-memory computing hardware architecture based on AM to support AMPER by leveraging parallel in-memory search operations. AMPER shows comparable learning performance while achieving 55x to 270x latency improvement when running on the proposed hardware compared to the state-of-the-art PER running on GPU.

In this paper, we propose Thinker, a bootstrapping method to remove adversarial effects of confounding features from the observation in an unsupervised way, and thus, it improves RL agents' generalization. Thinker first clusters experience trajectories into several clusters. These trajectories are then bootstrapped by applying a style transfer generator, which translates the trajectories from one cluster's style to another while maintaining the content of the observations. The bootstrapped trajectories are then used for policy learning. Thinker has wide applicability among many RL settings. Experimental results reveal that Thinker leads to better generalization capability in the Procgen benchmark environments compared to base algorithms and several data augmentation techniques.

Reinforcement learning (RL) workloads take a notoriously long time to train due to the large number of samples collected at run-time from simulators. Unfortunately, cluster scale-up approaches remain expensive, and commonly used CPU implementations of simulators induce high overhead when switching back and forth between GPU computations. We explore two optimizations that increase RL data collection efficiency by increasing GPU utilization: (1) GPU vectorization: parallelizing simulation on the GPU for increased hardware parallelism, and (2) simulator kernel fusion: fusing multiple simulation steps to run in a single GPU kernel launch to reduce global memory bandwidth requirements. We find that GPU vectorization can achieve up to $1024\times$ speedup over commonly used CPU simulators. We profile the performance of different implementations and show that for a simple simulator, ML compiler implementations (XLA) of GPU vectorization outperform a DNN framework (PyTorch) by $13.4\times$ by reducing CPU overhead from repeated Python to DL backend API calls. We show that simulator kernel fusion speedups with a simple simulator are $11.3\times$ and increase by up to $1024\times$ as simulator complexity increases in terms of memory bandwidth requirements. We show that the speedups from simulator kernel fusion are orthogonal and combinable with GPU vectorization, leading to a multiplicative speedup.

Machine Learning (ML) techniques have gained significant traction as a means of improving the autonomy of marine vehicles over the last few years. This article surveys the recent ML approaches utilised for ship collision avoidance (COLAV) and mission planning. Following an overview of the ever-expanding ML exploitation for maritime vehicles, key topics in the mission planning of ships are outlined. Notable papers with direct and indirect applications to the COLAV subject are technically reviewed and compared. Critiques, challenges, and future directions are also identified. The outcome clearly demonstrates the thriving research in this field, even though commercial marine ships incorporating machine intelligence able to perform autonomously under all operating conditions are still a long way off.

We present a novel generative method for producing unseen and plausible counterfactual examples for reinforcement learning (RL) agents based upon outcome variables that characterize agent behavior. Our approach uses a variational autoencoder to train a latent space that jointly encodes information about the observations and outcome variables pertaining to an agent's behavior. Counterfactuals are generated using traversals in this latent space, via gradient-driven updates as well as latent interpolations against cases drawn from a pool of examples. These include updates to raise the likelihood of generated examples, which improves the plausibility of generated counterfactuals. From experiments in three RL environments, we show that these methods produce counterfactuals that are more plausible and proximal to their queries compared to purely outcome-driven or case-based baselines. Finally, we show that a latent jointly trained to reconstruct both the input observations and behavioral outcome variables produces higher-quality counterfactuals over latents trained solely to reconstruct the observation inputs.

Asset allocation is an investment strategy that aims to balance risk and reward by constantly redistributing the portfolio's assets according to certain goals, risk tolerance, and investment horizon. Unfortunately, there is no simple formula that can find the right allocation for every individual. As a result, investors may use different asset allocations' strategy to try to fulfil their financial objectives. In this work, we conduct an extensive benchmark study to determine the efficacy and reliability of a number of optimization techniques. In particular, we focus on traditional approaches based on Modern Portfolio Theory, and on machine-learning approaches based on deep reinforcement learning. We assess the model's performance under different market tendency, i.e., both bullish and bearish markets. For reproducibility, we provide the code implementation code in this repository.

A realistic simulation environment is an essential tool in every roboticist's toolkit, with uses ranging from planning and control to training policies with reinforcement learning. Despite the centrality of simulation in modern robotics, little work has been done to compare the performance of robotics simulators against real-world data, especially for scenarios involving dynamic motions with high speed impact events. Handling dynamic contact is the computational bottleneck for most simulations, and thus the modeling and algorithmic choices surrounding impacts and friction form the largest distinctions between popular tools. Here, we evaluate the ability of several simulators to reproduce real-world trajectories involving impacts. Using experimental data, we identify system-specific contact parameters of popular simulators Drake, MuJoCo, and Bullet, analyzing the effects of modeling choices around these parameters. For the simple example of a cube tossed onto a table, simulators capture inelastic impacts well while failing to capture elastic impacts. For the higher-dimensional case of a Cassie biped landing from a jump, the simulators capture the bulk motion well but the accuracy is limited by numerous model differences between the real robot and the simulators.

Sepsis is a leading cause of death in the ICU. It is a disease requiring complex interventions in a short period of time, but its optimal treatment strategy remains uncertain. Evidence suggests that the practices of currently used treatment strategies are problematic and may cause harm to patients. To address this decision problem, we propose a new medical decision model based on historical data to help clinicians recommend the best reference option for real-time treatment. Our model combines offline reinforcement learning and deep reinforcement learning to solve the problem of traditional reinforcement learning in the medical field due to the inability to interact with the environment, while enabling our model to make decisions in a continuous state-action space. We demonstrate that, on average, the treatments recommended by the model are more valuable and reliable than those recommended by clinicians. In a large validation dataset, we find out that the patients whose actual doses from clinicians matched the decisions made by AI has the lowest mortality rates. Our model provides personalized and clinically interpretable treatment decisions for sepsis to improve patient care.

Single-machine and single-agent RL training have many challenges, the most important being the time it takes for the rewards to converge. Most of the time spent by the agent in RL training goes into gathering experiences. The time taken for simple applications is a few hours, and complex applications take days. Deep Learning frameworks like Tensorflow support distributed training; can the same be applied to RL as well? This article focuses on specific pain points of single-machine training with a practical example and demonstrates how scaled RL solves the problem.