We study offline reinforcement learning problems with a long-run average reward objective. The state-action pairs generated by any fixed behavioral policy thus follow a Markov chain, and the {\em empirical} state-action-next-state distribution satisfies a large deviations principle. We use the rate function of this large deviations principle to construct an uncertainty set for the unknown {\em true} state-action-next-state distribution. We also construct a distribution shift transformation that maps any distribution in this uncertainty set to a state-action-next-state distribution of the Markov chain generated by a fixed evaluation policy, which may differ from the unknown behavioral policy. We prove that the worst-case average reward of the evaluation policy with respect to all distributions in the shifted uncertainty set provides, in a rigorous statistical sense, the least conservative estimator for the average reward under the unknown true distribution. This guarantee is available even if one has only access to one single trajectory of serially correlated state-action pairs. The emerging robust optimization problem can be viewed as a robust Markov decision process with a non-rectangular uncertainty set. We adapt an efficient policy gradient algorithm to solve this problem. Numerical experiments show that our methods compare favorably against state-of-the-art methods.

Offline reinforcement learning (RL) looks at learning how to optimally solve tasks using a fixed dataset of interactions from the environment. Many off-policy algorithms developed for online learning struggle in the offline setting as they tend to over-estimate the behaviour of out of distributions actions. Existing offline RL algorithms adapt off-policy algorithms, employing techniques such as constraining the policy or modifying the value function to achieve good performance on individual datasets but struggle to adapt to different tasks or datasets of different qualities without tuning hyper-parameters. We introduce a policy switching technique that dynamically combines the behaviour of a pure off-policy RL agent, for improving behaviour, and a behavioural cloning (BC) agent, for staying close to the data. We achieve this by using a combination of epistemic uncertainty, quantified by our RL model, and a metric for aleatoric uncertainty extracted from the dataset. We show empirically that our policy switching technique can outperform not only the individual algorithms used in the switching process but also compete with state-of-the-art methods on numerous benchmarks. Our use of epistemic uncertainty for policy switching also allows us to naturally extend our method to the domain of offline to online fine-tuning allowing our model to adapt quickly and safely from online data, either matching or exceeding the performance of current methods that typically require additional modification or hyper-parameter fine-tuning.

High energy consumption remains a key challenge in high-performance computing (HPC) systems, which often feature hundreds or thousands of nodes drawing substantial power even in idle or standby modes. Although powering down unused nodes can improve energy efficiency, choosing the wrong time to do so can degrade quality of service by delaying job execution. Machine learning, in particular reinforcement learning (RL), has shown promise in determining optimal times to switch nodes on or off. In this study, we enhance the performance of a deep reinforcement learning (DRL) agent for HPC power management by integrating curriculum learning (CL), a training approach that introduces tasks with gradually increasing difficulty. Using the Batsim-py simulation framework, we compare the proposed CL-based agent to both a baseline DRL method (without CL) and the conventional fixed-time timeout strategy. Experimental results confirm that an easy-to-hard curriculum outperforms other training orders in terms of reducing wasted energy usage. The best agent achieves a 3.73% energy reduction over the baseline DRL method and a 4.66% improvement compared to the best timeout configuration (shutdown every 15 minutes of idle time). In addition, it reduces average job waiting time by 9.24% and maintains a higher job-filling rate, indicating more effective resource utilization. Sensitivity tests across various switch-on durations, power levels, and cluster sizes further reveal the agent's adaptability to changing system parameters without retraining. These findings demonstrate that curriculum learning can significantly improve DRL-based power management in HPC, balancing energy savings, quality of service, and robustness to diverse configurations.

Minimizing response times to meet legal requirements and serve patients in a timely manner is crucial for Emergency Medical Service (EMS) systems. Achieving this goal necessitates optimizing operational decision-making to efficiently manage ambulances. Against this background, we study a centrally controlled EMS system for which we learn an online ambulance dispatching and redeployment policy that aims at minimizing the mean response time of ambulances within the system by dispatching an ambulance upon receiving an emergency call and redeploying it to a waiting location upon the completion of its service. We propose a novel combinatorial optimization-augmented machine learning pipeline that allows to learn efficient policies for ambulance dispatching and redeployment. In this context, we further show how to solve the underlying full-information problem to generate training data and propose an augmentation scheme that improves our pipeline's generalization performance by mitigating a possible distribution mismatch with respect to the considered state space. Compared to existing methods that rely on augmentation during training, our approach offers substantial runtime savings of up to 87.9% while yielding competitive performance. To evaluate the performance of our pipeline against current industry practices, we conduct a numerical case study on the example of San Francisco's 911 call data. Results show that the learned policies outperform the online benchmarks across various resource and demand scenarios, yielding a reduction in mean response time of up to 30%.

Reinforcement Learning (RL) has proven largely effective in obtaining stable locomotion gaits for legged robots. However, designing control algorithms which can robustly navigate unseen environments with obstacles remains an ongoing problem within quadruped locomotion. To tackle this, it is convenient to solve navigation tasks by means of a hierarchical approach with a low-level locomotion policy and a high-level navigation policy. Crucially, the high-level policy needs to be robust to dynamic obstacles along the path of the agent. In this work, we propose a novel way to endow navigation policies with robustness by a training process that models obstacles as adversarial agents, following the adversarial RL paradigm. Importantly, to improve the reliability of the training process, we bound the rationality of the adversarial agent resorting to quantal response equilibria, and place a curriculum over its rationality. We called this method Hierarchical policies via Quantal response Adversarial Reinforcement Learning (Hi-QARL). We demonstrate the robustness of our method by benchmarking it in unseen randomized mazes with multiple obstacles. To prove its applicability in real scenarios, our method is applied on a Unitree GO1 robot in simulation.

Adaptive traffic signal control (ATSC) is crucial in reducing congestion, maximizing throughput, and improving mobility in rapidly growing urban areas. Recent advancements in parameter-sharing multi-agent reinforcement learning (MARL) have greatly enhanced the scalable and adaptive optimization of complex, dynamic flows in large-scale homogeneous networks. However, the inherent heterogeneity of real-world traffic networks, with their varied intersection topologies and interaction dynamics, poses substantial challenges to achieving scalable and effective ATSC across different traffic scenarios. To address these challenges, we present Unicorn, a universal and collaborative MARL framework designed for efficient and adaptable network-wide ATSC. Specifically, we first propose a unified approach to map the states and actions of intersections with varying topologies into a common structure based on traffic movements. Next, we design a Universal Traffic Representation (UTR) module with a decoder-only network for general feature extraction, enhancing the model's adaptability to diverse traffic scenarios. Additionally, we incorporate an Intersection Specifics Representation (ISR) module, designed to identify key latent vectors that represent the unique intersection's topology and traffic dynamics through variational inference techniques. To further refine these latent representations, we employ a contrastive learning approach in a self-supervised manner, which enables better differentiation of intersection-specific features. Moreover, we integrate the state-action dependencies of neighboring agents into policy optimization, which effectively captures dynamic agent interactions and facilitates efficient regional collaboration. Our results show that Unicorn outperforms other methods across various evaluation metrics, highlighting its potential in complex, dynamic traffic networks.

Diffusion models have achieved remarkable success in text-to-image generation. However, their practical applications are hindered by the misalignment between generated images and corresponding text prompts. To tackle this issue, reinforcement learning (RL) has been considered for diffusion model fine-tuning. Yet, RL's effectiveness is limited by the challenge of sparse reward, where feedback is only available at the end of the generation process. This makes it difficult to identify which actions during the denoising process contribute positively to the final generated image, potentially leading to ineffective or unnecessary denoising policies. To this end, this paper presents a novel RL-based framework that addresses the sparse reward problem when training diffusion models. Our framework, named $\text{B}^2\text{-DiffuRL}$, employs two strategies: \textbf{B}ackward progressive training and \textbf{B}ranch-based sampling. For one thing, backward progressive training focuses initially on the final timesteps of denoising process and gradually extends the training interval to earlier timesteps, easing the learning difficulty from sparse rewards. For another, we perform branch-based sampling for each training interval. By comparing the samples within the same branch, we can identify how much the policies of the current training interval contribute to the final image, which helps to learn effective policies instead of unnecessary ones. $\text{B}^2\text{-DiffuRL}$ is compatible with existing optimization algorithms. Extensive experiments demonstrate the effectiveness of $\text{B}^2\text{-DiffuRL}$ in improving prompt-image alignment and maintaining diversity in generated images. The code for this work is available.

It is commonly believed that, in a real-world environment, samples can only be drawn from observational and interventional distributions, corresponding to Layers 1 and 2 of the Pearl Causal Hierarchy. Layer 3, representing counterfactual distributions, is believed to be inaccessible by definition. However, Bareinboim, Forney, and Pearl (2015) introduced a procedure that allows an agent to sample directly from a counterfactual distribution, leaving open the question of what other counterfactual quantities can be estimated directly via physical experimentation. We resolve this by introducing a formal definition of realizability, the ability to draw samples from a distribution, and then developing a complete algorithm to determine whether an arbitrary counterfactual distribution is realizable given fundamental physical constraints, such as the inability to go back in time and subject the same unit to a different experimental condition. We illustrate the implications of this new framework for counterfactual data collection using motivating examples from causal fairness and causal reinforcement learning. While the baseline approach in these motivating settings typically follows an interventional or observational strategy, we show that a counterfactual strategy provably dominates both.

Spasticity is a common movement disorder symptom in individuals with cerebral palsy, hereditary spastic paraplegia, spinal cord injury and stroke, being one of the most disabling features in the progression of these diseases. Despite the potential benefit of using wearable robots to treat spasticity, their use is not currently recommended to subjects with a level of spasticity above ${1^+}$ on the Modified Ashworth Scale. The varying dynamics of this velocity-dependent tonic stretch reflex make it difficult to deploy safe personalized controllers. Here, we describe a novel adaptive torque controller via deep reinforcement learning (RL) for a knee exoskeleton under joint spasticity conditions, which accounts for task performance and interaction forces reduction. To train the RL agent, we developed a digital twin, including a musculoskeletal-exoskeleton system with joint misalignment and a differentiable spastic reflexes model for the muscles activation. Results for a simulated knee extension movement showed that the agent learns to control the exoskeleton for individuals with different levels of spasticity. The proposed controller was able to reduce maximum torques applied to the human joint under spastic conditions by an average of 10.6\% and decreases the root mean square until the settling time by 8.9\% compared to a conventional compliant controller.

Recent research on Reasoning of Large Language Models (LLMs) has sought to further enhance their performance by integrating meta-thinking -- enabling models to monitor, evaluate, and control their reasoning processes for more adaptive and effective problem-solving. However, current single-agent work lacks a specialized design for acquiring meta-thinking, resulting in low efficacy. To address this challenge, we introduce Reinforced Meta-thinking Agents (ReMA), a novel framework that leverages Multi-Agent Reinforcement Learning (MARL) to elicit meta-thinking behaviors, encouraging LLMs to think about thinking. ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions. Through iterative reinforcement learning with aligned objectives, these agents explore and learn collaboration, leading to improved generalization and robustness. Experimental results demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks, including competitive-level mathematical benchmarks and LLM-as-a-Judge benchmarks. Comprehensive ablation studies further illustrate the evolving dynamics of each distinct agent, providing valuable insights into how the meta-thinking reasoning process enhances the reasoning capabilities of LLMs.