We evaluate BARL on the TQRL setting in 5 environments which span a variety of reward function types, dimensionalities, and amounts of required data. In this evaluation, we estimate the minimum amount of data an algorithm needs to learn a controller. The evaluation environments include the standard underactuated pendulum swing-up task, a cartpole swing-up task, the standard 2-DOF reacher task, a navigation problem where the agent must find a path across pools of lava, and a simulated nuclear fusion control problem where the agent is tasked with modulating the power injected into the plasma to achieve a target pressure. To assess the performance of BARL in solving MDPs quickly, we assembled a group of reinforcement learning algorithms that represent the state of the art in solving continuous MDPs. We compare against model-based algorithms PILCO [7], PETS [2], model-predictive control with a GP (MPC), and uncertainty sampling with a GP (), as well as model-free algorithms SAC [3], TD3 [8], and PPO [9].

Figure 1: Summary of our recommendations for when a practitioner should BC and various imitation learning style methods, and when they should use offline RL approaches. Offline reinforcement learning allows learning policies from previously collected data, which has profound implications for applying RL in domains where running trial-and-error learning is impractical or dangerous, such as safety-critical settings like autonomous driving or medical treatment planning. In such scenarios, online exploration is simply too risky, but offline RL methods can learn effective policies from logged data collected by humans or heuristically designed controllers. Prior learning-based control methods have also approached learning from existing data as imitation learning: if the data is generally "good enough," simply copying the behavior in the data can lead to good results, and if it's not good enough, then filtering or reweighting the data and then copying can work well. Several recent works suggest that this is a viable alternative to modern offline RL methods.

Prediction intervals offer an effective tool for quantifying the uncertainty of loads in distribution systems. The traditional central PIs cannot adapt well to skewed distributions, and their offline training fashion is vulnerable to unforeseen changes in future load patterns. Therefore, we propose an optimal PI estimation approach, which is online and adaptive to different data distributions by adaptively determining symmetric or asymmetric probability proportion pairs for quantiles. It relies on the online learning ability of reinforcement learning to integrate the two online tasks, i.e., the adaptive selection of probability proportion pairs and quantile predictions, both of which are modeled by neural networks. As such, the quality of quantiles-formed PI can guide the selection process of optimal probability proportion pairs, which forms a closed loop to improve the quality of PIs. Furthermore, to improve the learning efficiency of quantile forecasts, a prioritized experience replay strategy is proposed for online quantile regression processes. Case studies on both load and net load demonstrate that the proposed method can better adapt to data distribution compared with online central PIs method. Compared with offline-trained methods, it obtains PIs with better quality and is more robust against concept drift.

This work identifies a common flaw of deep reinforcement learning (RL) algorithms: a tendency to rely on early interactions and ignore useful evidence encountered later. Because of training on progressively growing datasets, deep RL agents incur a risk of overfitting to earlier experiences, negatively affecting the rest of the learning process. Inspired by cognitive science, we refer to this effect as the primacy bias. Through a series of experiments, we dissect the algorithmic aspects of deep RL that exacerbate this bias. We then propose a simple yet generally-applicable mechanism that tackles the primacy bias by periodically resetting a part of the agent. We apply this mechanism to algorithms in both discrete (Atari 100k) and continuous action (DeepMind Control Suite) domains, consistently improving their performance.

Deep reinforcement learning (DRL) is transitioning from a research field focused on game playing to a technology with real-world applications. Notable examples include DeepMind's work on controlling a nuclear reactor or on improving Youtube video compression, or Tesla attempting to use a method inspired by MuZero for autonomous vehicle behavior planning. But the exciting potential for real world applications of RL should also come with a healthy dose of caution – for example RL policies are well known to be vulnerable to exploitation, and methods for safe and robust policy development are an active area of research. At the same time as the emergence of powerful RL systems in the real world, the public and researchers are expressing an increased appetite for fair, aligned, and safe machine learning systems. The focus of these research efforts to date has been to account for shortcomings of datasets or supervised learning practices that can harm individuals.

This paper presents a novel trajectory tracker for autonomous quadrotor navigation in dynamic and complex environments. The proposed framework integrates a distributional Reinforcement Learning (RL) estimator for unknown aerodynamic effects into a Stochastic Model Predictive Controller (SMPC) for trajectory tracking. Aerodynamic effects derived from drag forces and moment variations are difficult to model directly and accurately. Most current quadrotor tracking systems therefore treat them as simple `disturbances' in conventional control approaches. We propose Quantile-approximation-based Distributional Reinforced-disturbance-estimator, an aerodynamic disturbance estimator, to accurately identify disturbances, i.e., uncertainties between the true and estimated values of aerodynamic effects. Simplified Affine Disturbance Feedback is employed for control parameterization to guarantee convexity, which we then integrate with a SMPC to achieve sufficient and non-conservative control signals. We demonstrate our system to improve the cumulative tracking errors by at least 66% with unknown and diverse aerodynamic forces compared with recent state-of-the-art. Concerning traditional Reinforcement Learning's non-interpretability, we provide convergence and stability guarantees of Distributional RL and SMPC, respectively, with non-zero mean disturbances.

In emerging Industrial Cyber-Physical Systems (ICPSs), the joint design of communication and control sub-systems is essential, as these sub-systems are interconnected. In this paper, we study the joint design problem of an event-triggered control and an energy-efficient resource allocation in a fifth generation (5G) wireless network. We formally state the problem as a multi-objective optimization one, aiming to minimize the number of updates on the actuators' input and the power consumption in the downlink transmission. To address the problem, we propose a model-free hierarchical reinforcement learning approach \textcolor{blue}{with uniformly ultimate boundedness stability guarantee} that learns four policies simultaneously. These policies contain an update time policy on the actuators' input, a control policy, and energy-efficient sub-carrier and power allocation policies. Our simulation results show that the proposed approach can properly control a simulated ICPS and significantly decrease the number of updates on the actuators' input as well as the downlink power consumption.

Fast and reliable connectivity is essential to enhancing situational awareness and operational efficiency for public safety mission-critical (MC) users. In emergency or disaster circumstances, where existing cellular network coverage and capacity may not be available to meet MC communication demands, deployable-network-based solutions such as cells-on-wheels/wings can be utilized swiftly to ensure reliable connection for MC users. In this paper, we consider a scenario where a macro base station (BS) is destroyed due to a natural disaster and an unmanned aerial vehicle carrying BS (UAV-BS) is set up to provide temporary coverage for users in the disaster area. The UAV-BS is integrated into the mobile network using the 5G integrated access and backhaul (IAB) technology. We propose a framework and signalling procedure for applying machine learning to this use case. A deep reinforcement learning algorithm is designed to jointly optimize the access and backhaul antenna tilt as well as the three-dimensional location of the UAV-BS in order to best serve the on-ground MC users while maintaining a good backhaul connection. Our result shows that the proposed algorithm can autonomously navigate and configure the UAV-BS to improve the throughput and reduce the drop rate of MC users.

Upside-Down Reinforcement Learning (UDRL) is an approach for solving RL problems that does not require value functions and uses only supervised learning, where the targets for given inputs in a dataset do not change over time [4, 5]. Ghosh et al. [2] proved that Goal-Conditional Supervised Learning (GCSL)--which can be viewed as a simplified version of UDRL--optimizes a lower bound on goal-reaching performance. This raises expectations that such algorithms may enjoy guaranteed convergence to the optimal policy in arbitrary environments, similar to certain well-known traditional RL algorithms. Here we show that for a specific episodic UDRL algorithm (eUDRL, including GCSL), this is not the case, and give the causes of this limitation. To do so, we first introduce a helpful rewrite of eUDRL as a recursive policy update. This formulation helps to disprove its convergence to the optimal policy for a wide class of stochastic environments. Finally, we provide a concrete example of a very simple environment where eUDRL diverges. Since the primary aim of this paper is to present a negative result, and the best counterexamples are the simplest ones, we restrict all discussions to finite (discrete) environments, ignoring issues of function approximation and limited sample size.

Reinforcement learning (RL) is highly suitable for devising control strategies in the context of dynamical systems. A prominent instance of such a dynamical system is the system of equations governing fluid dynamics. Recent research results indicate that RL-augmented computational fluid dynamics (CFD) solvers can exceed the current state of the art, for example in the field of turbulence modeling. However, while in supervised learning, the training data can be generated a priori in an offline manner, RL requires constant run-time interaction and data exchange with the CFD solver during training. In order to leverage the potential of RL-enhanced CFD, the interaction between the CFD solver and the RL algorithm thus have to be implemented efficiently on high-performance computing (HPC) hardware. To this end, we present Relexi as a scalable RL framework that bridges the gap between machine learning workflows and modern CFD solvers on HPC systems providing both components with its specialized hardware. Relexi is built with modularity in mind and allows easy integration of various HPC solvers by means of the in-memory data transfer provided by the SmartSim library. Here, we demonstrate that the Relexi framework can scale up to hundreds of parallel environment on thousands of cores. This allows to leverage modern HPC resources to either enable larger problems or faster turnaround times. Finally, we demonstrate the potential of an RL-augmented CFD solver by finding a control strategy for optimal eddy viscosity selection in large eddy simulations.