We propose VRL3, a powerful data-driven framework with a simple design for solving challenging visual deep reinforcement learning (DRL) tasks. We analyze a number of major obstacles in taking a data-driven approach, and present a suite of design principles, novel findings, and critical insights about data-driven visual DRL. Our framework has three stages: in stage 1, we leverage non-RL datasets (e.g. ImageNet) to learn task-agnostic visual representations; in stage 2, we use offline RL data (e.g. a limited number of expert demonstrations) to convert the task-agnostic representations into more powerful task-specific representations; in stage 3, we fine-tune the agent with online RL. On a set of challenging hand manipulation tasks with sparse reward and realistic visual inputs, compared to the previous SOTA, VRL3 achieves an average of 780% better sample efficiency.

This paper presents a comprehensive system model for goodput maximization with quantized feedback in Ultra-Reliable Low-Latency Communication (URLLC), focusing on dynamic channel conditions and feedback schemes. The study investigates a communication system, where the receiver provides quantized channel state information to the transmitter. The system adapts its feedback scheme based on reinforcement learning, aiming to maximize goodput while accommodating varying channel statistics. We introduce a novel Rician-$K$ factor estimation technique to enable the communication system to optimize the feedback scheme. This dynamic approach increases the overall performance, making it well-suited for practical URLLC applications where channel statistics vary over time.

Attention can be used to inform choice selection in contextual bandit tasks even when context features have not been previously experienced. One example of this is in dimensional shifts, where additional feature values are introduced and the relationship between features and outcomes can either be static or variable. Attentional mechanisms have been extensively studied in contextual bandit tasks where the feedback of choices is provided immediately, but less research has been done on tasks where feedback is delayed or in counterfactual feedback cases. Some methods have successfully modeled human attention with immediate feedback based on reward prediction errors (RPEs), though recent research raises questions of the applicability of RPEs onto more general attentional mechanisms. Alternative models suggest that information theoretic metrics can be used to model human attention, with broader applications to novel stimuli. In this paper, we compare two different methods for modeling how humans attend to specific features of decision making tasks, one that is based on calculating an information theoretic metric using a memory of past experiences, and another that is based on iteratively updating attention from reward prediction errors. We compare these models using simulations in a contextual bandit task with both intradimensional and extradimensional domain shifts, as well as immediate, delayed, and counterfactual feedback. We find that calculating an information theoretic metric over a history of experiences is best able to account for human-like behavior in tasks that shift dimensions and alter feedback presentation. These results indicate that information theoretic metrics of attentional mechanisms may be better suited than RPEs to predict human attention in decision making, though further studies of human behavior are necessary to support these results.

In this paper, a novel network architecture that deploys the multi-functional reconfigurable intelligent surface (MF-RIS) in low-Earth orbit (LEO) is proposed. Unlike traditional RIS with only signal reflection capability, the MF-RIS can reflect, refract, and amplify signals, as well as harvest energy from wireless signals. Given the high energy demands in shadow regions where solar energy is unavailable, MF-RIS is deployed in LEO to enhance signal coverage and improve energy efficiency (EE). To address this, we formulate a long-term EE optimization problem by determining the optimal parameters for MF-RIS configurations, including amplification and phase-shifts, energy harvesting ratios, and LEO transmit beamforming. To address the complex non-convex and non-linear problem, a federated learning enhanced multi-agent deep deterministic policy gradient (FEMAD) scheme is designed. Multi-agent DDPG of each agent can provide the optimal action policy from its interaction to environments, whereas federated learning enables the hidden information exchange among multi-agents. In numerical results, we can observe significant EE improvements compared to the other benchmarks, including centralized deep reinforcement learning as well as distributed multi-agent deep deterministic policy gradient (DDPG). Additionally, the proposed LEO-MF-RIS architecture has demonstrated its effectiveness, achieving the highest EE performance compared to the scenarios of fixed/no energy harvesting in MF-RIS, traditional reflection-only RIS, and deployment without RISs/MF-RISs.

The Configurable Markov Decision Process framework includes two entities: a Reinforcement Learning agent and a configurator that can modify some environmental parameters to improve the agent's performance. This presupposes that the two actors have the same reward functions. What if the configurator does not have the same intentions as the agent? This paper introduces the Non-Cooperative Configurable Markov Decision Process, a setting that allows having two (possibly different) reward functions for the configurator and the agent. Then, we consider an online learning problem, where the configurator has to find the best among a finite set of possible configurations.

Constructing Reinforcement Learning (RL) policies that adhere to safety requirements is an emerging field of study. RL agents learn via trial and error with an objective to optimize a reward signal. Often policies that are designed to accumulate rewards do not satisfy safety specifications. We present a methodology for counterexample guided refinement of a trained RL policy against a given safety specification. Our approach has two main components.

In many real-world scenarios, Reinforcement Learning (RL) algorithms are trained on data with dynamics shift, i.e., with different underlying environment dynamics. A majority of current methods address such issue by training context encoders to identify environment parameters. Data with dynamics shift are separated according to their environment parameters to train the corresponding policy.However, these methods can be sample inefficient as data are used \textit{ad hoc}, and policies trained for one dynamics cannot benefit from data collected in all other environments with different dynamics. In this paper, we find that in many environments with similar structures and different dynamics, optimal policies have similar stationary state distributions. We exploit such property and learn the stationary state distribution from data with dynamics shift for efficient data reuse.

We study adapting trained object detectors to unseen domains manifesting significant variations of object appearance, viewpoints and backgrounds. Most current methods align domains by either using image or instance-level feature alignment in an adversarial fashion. This often suffers due to the presence of unwanted background and as such lacks class-specific alignment. A common remedy to promote class-level alignment is to use high confidence predictions on the unlabelled domain as pseudo labels. These high confidence predictions are often fallacious since the model is poorly calibrated under domain shift.

We extend the options framework for temporal abstraction in reinforcement learning from discounted Markov decision processes (MDPs) to average-reward MDPs. Our contributions include general convergent off-policy inter-option learning algorithms, intra-option algorithms for learning values and models, as well as sample-based planning variants of our learning algorithms. Our algorithms and convergence proofs extend those recently developed by Wan, Naik, and Sutton. We also extend the notion of option-interrupting behaviour from the discounted to the average-reward formulation. We show the efficacy of the proposed algorithms with experiments on a continuing version of the Four-Room domain.

In risk-averse reinforcement learning (RL), the goal is to optimize some risk measure of the returns. A risk measure often focuses on the worst returns out of the agent's experience. As a result, standard methods for risk-averse RL often ignore high-return strategies. We prove that under certain conditions this inevitably leads to a local-optimum barrier, and propose a mechanism we call soft risk to bypass it. We also devise a novel cross entropy module for sampling, which (1) preserves risk aversion despite the soft risk; (2) independently improves sample efficiency.