Real-world reinforcement learning (RL) environments, whether in robotics or industrial settings, often involve non-visual observations and require not only efficient but also reliable and thus interpretable and flexible RL approaches. To improve efficiency, agents that perform state representation learning with auxiliary tasks have been widely studied in visual observation contexts. However, for real-world problems, dedicated representation learning modules that are decoupled from RL agents are more suited to meet requirements. This study compares common auxiliary tasks based on, to the best of our knowledge, the only decoupled representation learning method for low-dimensional non-visual observations. We evaluate potential improvements in sample efficiency and returns for environments ranging from a simple pendulum to a complex simulated robotics task. Our findings show that representation learning with auxiliary tasks only provides performance gains in sufficiently complex environments and that learning environment dynamics is preferable to predicting rewards. These insights can inform future development of interpretable representation learning approaches for non-visual observations and advance the use of RL solutions in real-world scenarios.

Various mobile applications that comprise dependent tasks are gaining widespread popularity and are increasingly complex. These applications often have low-latency requirements, resulting in a significant surge in demand for computing resources. With the emergence of mobile edge computing (MEC), it becomes the most significant issue to offload the application tasks onto small-scale devices deployed at the edge of the mobile network for obtaining a high-quality user experience. However, since the environment of MEC is dynamic, most existing works focusing on task graph offloading, which rely heavily on expert knowledge or accurate analytical models, fail to fully adapt to such environmental changes, resulting in the reduction of user experience. This paper investigates the task graph offloading in MEC, considering the time-varying computation capabilities of edge computing devices. To adapt to environmental changes, we model the task graph scheduling for computation offloading as a Markov Decision Process (MDP). Then, we design a deep reinforcement learning algorithm (SATA-DRL) to learn the task scheduling strategy from the interaction with the environment, to improve user experience. Extensive simulations validate that SATA-DRL is superior to existing strategies in terms of reducing average makespan and deadline violation.

Aligning language models (LMs) with preferences is an important problem in natural language generation. A key challenge is that preferences are typically provided at the *sequence level* while LM training and generation both occur at the *token level*. There is, therefore, a *granularity mismatch* between the preference and the LM training losses, which may complicate the learning problem. In this paper, we address this issue by developing an alternate training process, where we iterate between grounding the sequence-level preference into token-level training guidance, and improving the LM with the learned guidance. For guidance learning, we design a framework that extends the pairwise-preference learning in imitation learning to both variable-length LM generation and the utilization of the preference among multiple generations. For LM training, based on the amount of supervised data, we present two *minimalist* learning objectives that utilize the learned guidance. In experiments, our method performs competitively on two distinct representative LM tasks -- discrete-prompt generation and text summarization.

Algorithmic decision-making in practice must be fair for legal, ethical, and societal reasons. To achieve this, prior research has contributed various approaches that ensure fairness in machine learning predictions, while comparatively little effort has focused on fairness in decision-making, specifically off-policy learning. In this paper, we propose a novel framework for fair off-policy learning: we learn decision rules from observational data under different notions of fairness, where we explicitly assume that observational data were collected under a different - potentially discriminatory - behavioral policy. We then propose a neural networkbased framework to learn optimal policies under different fairness notions. We further provide theoretical guarantees in the form of generalization bounds for the finite-sample version of our framework. We demonstrate the effectiveness of our framework through extensive numerical experiments using both simulated and real-world data. Altogether, our work enables algorithmic decision-making in a wide array of practical applications where fairness must be ensured. Algorithmic decision-making in practice must avoid discrimination and thus be fair to meet legal, ethical, and societal demands (Nkonde, 2019; De-Arteaga et al., 2022; Corbett-Davies et al., 2023). For example, in the U.S., the Fair Housing Act and Equal Credit Opportunity Act stipulate that decisions must not be subject to systematic discrimination by gender, race, or other attributes deemed as sensitive. However, research from different areas has provided repeated evidence that algorithmic decisionmaking is often not fair. A prominent example is Amazon's tool for automatically screening job applicants that was used between 2014 and 2017 (Dastin, 2018). It was later discovered that the underlying algorithm generated decisions that were subject to systematic discrimination against women and thus resulted in a ceteris paribus lower probability of women being hired. Ensuring fairness in off-policy learning is subject to inherent challenges.

We propose A-Crab (Actor-Critic Regularized by Average Bellman error), a new practical algorithm for offline reinforcement learning (RL) in complex environments with insufficient data coverage. Our algorithm combines the marginalized importance sampling framework with the actor-critic paradigm, where the critic returns evaluations of the actor (policy) that are pessimistic relative to the offline data and have a small average (importance-weighted) Bellman error. Compared to existing methods, our algorithm simultaneously offers a number of advantages: (1) It achieves the optimal statistical rate of $1/\sqrt{N}$ -- where $N$ is the size of offline dataset -- in converging to the best policy covered in the offline dataset, even when combined with general function approximators. (2) It relies on a weaker average notion of policy coverage (compared to the $\ell_\infty$ single-policy concentrability) that exploits the structure of policy visitations. (3) It outperforms the data-collection behavior policy over a wide range of specific hyperparameters. We provide both theoretical analysis and experimental results to validate the effectiveness of our proposed algorithm.

Reinforcement learning (RL) agents have long sought to approach the efficiency of human learning. Humans are great observers who can learn by aggregating external knowledge from various sources, including observations from others' policies of attempting a task. Prior studies in RL have incorporated external knowledge policies to help agents improve sample efficiency. However, it remains non-trivial to perform arbitrary combinations and replacements of those policies, an essential feature for generalization and transferability. In this work, we present Knowledge-Grounded RL (KGRL), an RL paradigm fusing multiple knowledge policies and aiming for human-like efficiency and flexibility. We propose a new actor architecture for KGRL, Knowledge-Inclusive Attention Network (KIAN), which allows free knowledge rearrangement due to embedding-based attentive action prediction. KIAN also addresses entropy imbalance, a problem arising in maximum entropy KGRL that hinders an agent from efficiently exploring the environment, through a new design of policy distributions. The experimental results demonstrate that KIAN outperforms alternative methods incorporating external knowledge policies and achieves efficient and flexible learning. Our implementation is available at https://github.com/Pascalson/KGRL.git

We study the problem of agnostic PAC reinforcement learning (RL): given a policy class $\Pi$, how many rounds of interaction with an unknown MDP (with a potentially large state and action space) are required to learn an $\epsilon$-suboptimal policy with respect to $\Pi$? Towards that end, we introduce a new complexity measure, called the \emph{spanning capacity}, that depends solely on the set $\Pi$ and is independent of the MDP dynamics. With a generative model, we show that for any policy class $\Pi$, bounded spanning capacity characterizes PAC learnability. However, for online RL, the situation is more subtle. We show there exists a policy class $\Pi$ with a bounded spanning capacity that requires a superpolynomial number of samples to learn. This reveals a surprising separation for agnostic learnability between generative access and online access models (as well as between deterministic/stochastic MDPs under online access). On the positive side, we identify an additional \emph{sunflower} structure, which in conjunction with bounded spanning capacity enables statistically efficient online RL via a new algorithm called POPLER, which takes inspiration from classical importance sampling methods as well as techniques for reachable-state identification and policy evaluation in reward-free exploration.

A model among many may only be best under certain states of the world. Switching from a model to another can also be costly. Finding a procedure to dynamically choose a model in these circumstances requires to solve a complex estimation procedure and a dynamic programming problem. A Reinforcement learning algorithm is used to approximate and estimate from the data the optimal solution to this dynamic programming problem. The algorithm is shown to consistently estimate the optimal policy that may choose different models based on a set of covariates. A typical example is the one of switching between different portfolio models under rebalancing costs, using macroeconomic information. Using a set of macroeconomic variables and price data, an empirical application to the aforementioned portfolio problem shows superior performance to choosing the best portfolio model with hindsight.

Deriving robust control policies for realistic urban navigation scenarios is not a trivial task. In an end-to-end approach, these policies must map high-dimensional images from the vehicle's cameras to low-level actions such as steering and throttle. While pure Reinforcement Learning (RL) approaches are based exclusively on rewards,Generative Adversarial Imitation Learning (GAIL) agents learn from expert demonstrations while interacting with the environment, which favors GAIL on tasks for which a reward signal is difficult to derive. In this work, the hGAIL architecture was proposed to solve the autonomous navigation of a vehicle in an end-to-end approach, mapping sensory perceptions directly to low-level actions, while simultaneously learning mid-level input representations of the agent's environment. The proposed hGAIL consists of an hierarchical Adversarial Imitation Learning architecture composed of two main modules: the GAN (Generative Adversarial Nets) which generates the Bird's-Eye View (BEV) representation mainly from the images of three frontal cameras of the vehicle, and the GAIL which learns to control the vehicle based mainly on the BEV predictions from the GAN as input.Our experiments have shown that GAIL exclusively from cameras (without BEV) fails to even learn the task, while hGAIL, after training, was able to autonomously navigate successfully in all intersections of the city.

Multi-armed bandits (MAB) (Auer, 2002) is a classic framework of sequential decision-making problems that has been extensively studied (Lattimore & Szepesvári, 2020; Slivkins et al., 2019). In each round, the learning agent selects one out of m arms and observes its reward feedback which follows an unknown reward distribution. The goal is to maximize the cumulative reward, which requires the agent to balance exploitation (selecting the arm with the highest average reward) and exploration (exploring arms that have high potential but have not been played enough). Combinatorial multi-armed bandits (CMAB) is a generalized setting of original MAB with many real-world applications such as online advertising, ranking, and influence maximization (Liu & Zhao, 2012; Kveton et al., 2015; Chen et al., 2016; Wang & Chen, 2017). In CMAB, the agent chooses a combinatorial action (called a super arm) over the m base arms in each round, and observes outcomes of base arms triggered by the action as feedback, known as the semi-bandit feedback.