Many applications in Reinforcement Learning (RL) usually have noise or stochasticity present in the environment. Beyond their impact on learning, these uncertainties lead the exact same policy to perform differently, i.e. yield different return, from one roll-out to another. Common evaluation procedures in RL summarise the consequent return distributions using solely the expected return, which does not account for the spread of the distribution. Our work defines this spread as the policy reproducibility: the ability of a policy to obtain similar performance when rolled out many times, a crucial property in some real-world applications. We highlight that existing procedures that only use the expected return are limited on two fronts: first an infinite number of return distributions with a wide range of performance-reproducibility trade-offs can have the same expected return, limiting its effectiveness when used for comparing policies; second, the expected return metric does not leave any room for practitioners to choose the best trade-off value for considered applications. In this work, we address these limitations by recommending the use of Lower Confidence Bound, a metric taken from Bayesian optimisation that provides the user with a preference parameter to choose a desired performance-reproducibility trade-off. We also formalise and quantify policy reproducibility, and demonstrate the benefit of our metrics using extensive experiments of popular RL algorithms on common uncertain RL tasks.

Recent studies have uncovered the potential of Large Language Models (LLMs) in addressing complex sequential decision-making tasks through the provision of high-level instructions. However, LLM-based agents lack specialization in tackling specific target problems, particularly in real-time dynamic environments. Additionally, deploying an LLM-based agent in practical scenarios can be both costly and time-consuming. On the other hand, reinforcement learning (RL) approaches train agents that specialize in the target task but often suffer from low sampling efficiency and high exploration costs. In this paper, we introduce a novel framework that addresses these challenges by training a smaller, specialized student RL agent using instructions from an LLM-based teacher agent. By incorporating the guidance from the teacher agent, the student agent can distill the prior knowledge of the LLM into its own model. Consequently, the student agent can be trained with significantly less data. Moreover, through further training with environment feedback, the student agent surpasses the capabilities of its teacher for completing the target task. We conducted experiments on challenging MiniGrid and Habitat environments, specifically designed for embodied AI research, to evaluate the effectiveness of our framework. The results clearly demonstrate that our approach achieves superior performance compared to strong baseline methods. Our code is available at https://github.com/ZJLAB-AMMI/LLM4Teach.

Biomimetic, dexterous robotic hands have the potential to replicate much of the tasks that a human can do, and to achieve status as a general manipulation platform. Recent advances in reinforcement learning (RL) frameworks have achieved remarkable performance in quadrupedal locomotion and dexterous manipulation tasks. Combined with GPU-based highly parallelized simulations capable of simulating thousands of robots in parallel, RL-based controllers have become more scalable and approachable. However, in order to bring RL-trained policies to the real world, we require training frameworks that output policies that can work with physical actuators and sensors as well as a hardware platform that can be manufactured with accessible materials yet is robust enough to run interactive policies. This work introduces the biomimetic tendon-driven Faive Hand and its system architecture, which uses tendon-driven rolling contact joints to achieve a 3D printable, robust high-DoF hand design. We model each element of the hand and integrate it into a GPU simulation environment to train a policy with RL, and achieve zero-shot transfer of a dexterous in-hand sphere rotation skill to the physical robot hand.

This Temporal-difference (TD) learning [1] is one of the most popular approach does not allow general and formal analysis frameworks reinforcement learning (RL) algorithms [2] for policy evaluation because the asymptotic stability of the ODE model significantly problems. However, its main limitation lies in its inability to accommodate depends on the specific algorithm, and it is in general hard to both off-policy learning and linear function approximation for establish the stability of the ODE model. On the other hand, convergence guarantees, which has been an important open problem the proposed analysis applies the recent asymptotic stability for decades. In 2009, Sutton, Szepesvári, and Maei [3], [4] introduced theory of primal-dual gradient dynamics (PDGD) [13], where the first TD learning algorithms compatible with both linear function control theoretic frameworks for stability analysis of PDGD are approximation and off-policy training based on gradient estimations, developed. Using this recent result, we provide a new template which are thus called gradient temporal-difference learning (GTD).

In software-defined networking (SDN), the implementation of distributed SDN controllers, with each controller responsible for managing a specific sub-network or domain, plays a critical role in achieving a balance between centralized control, scalability, reliability, and network efficiency. These controllers must be synchronized to maintain a logically centralized view of the entire network. While there are various approaches for synchronizing distributed SDN controllers, most tend to prioritize goals such as optimization of communication latency or load balancing, often neglecting to address both the aspects simultaneously. This limitation becomes particularly significant when considering applications like Augmented and Virtual Reality (AR/VR), which demand constrained network latencies and substantial computational resources. Additionally, many existing studies in this field predominantly rely on value-based reinforcement learning (RL) methods, overlooking the potential advantages offered by state-of-the-art policy-based RL algorithms. To bridge this gap, our work focuses on examining deep reinforcement learning (DRL) techniques, encompassing both value-based and policy-based methods, to guarantee an upper latency threshold for AR/VR task offloading within SDN environments, while selecting the most cost-effective servers for AR/VR task offloading. Our evaluation results indicate that while value-based methods excel in optimizing individual network metrics such as latency or load balancing, policy-based approaches exhibit greater robustness in adapting to sudden network changes or reconfiguration.

In upcoming 6G networks, unmanned aerial vehicles (UAVs) are expected to play a fundamental role by acting as mobile base stations, particularly for demanding vehicle-to-everything (V2X) applications. In this scenario, one of the most challenging problems is the design of trajectories for multiple UAVs, cooperatively serving the same area. Such joint trajectory design can be performed using multi-agent deep reinforcement learning (MADRL) algorithms, but ensuring collision-free paths among UAVs becomes a critical challenge. Traditional methods involve imposing high penalties during training to discourage unsafe conditions, but these can be proven to be ineffective, whereas binary masks can be used to restrict unsafe actions, but naively applying them to all agents can lead to suboptimal solutions and inefficiencies. To address these issues, we propose a rank-based binary masking approach. Higher-ranked UAVs move optimally, while lower-ranked UAVs use this information to define improved binary masks, reducing the number of unsafe actions. This approach allows to obtain a good trade-off between exploration and exploitation, resulting in enhanced training performance, while maintaining safety constraints.

Catastrophic forgetting poses a substantial challenge for managing intelligent agents controlled by a large model, causing performance degradation when these agents face new tasks. In our work, we propose a novel solution - the Progressive Prompt Decision Transformer (P2DT). This method enhances a transformer-based model by dynamically appending decision tokens during new task training, thus fostering task-specific policies. Our approach mitigates forgetting in continual and offline reinforcement learning scenarios. Moreover, P2DT leverages trajectories collected via traditional reinforcement learning from all tasks and generates new task-specific tokens during training, thereby retaining knowledge from previous studies. Preliminary results demonstrate that our model effectively alleviates catastrophic forgetting and scales well with increasing task environments.

Approximate Thompson sampling with Langevin Monte Carlo broadens its reach from Gaussian posterior sampling to encompass more general smooth posteriors. However, it still encounters scalability issues in high-dimensional problems when demanding high accuracy. To address this, we propose an approximate Thompson sampling strategy, utilizing underdamped Langevin Monte Carlo, where the latter is the go-to workhorse for simulations of high-dimensional posteriors. Based on the standard smoothness and log-concavity conditions, we study the accelerated posterior concentration and sampling using a specific potential function. This design improves the sample complexity for realizing logarithmic regrets from $\mathcal{\tilde O}(d)$ to $\mathcal{\tilde O}(\sqrt{d})$. The scalability and robustness of our algorithm are also empirically validated through synthetic experiments in high-dimensional bandit problems.

The study proposes the reformulation of offline reinforcement learning as a regression problem that can be solved with decision trees. Aiming to predict actions based on input states, return-to-go (RTG), and timestep information, we observe that with gradient-boosted trees, the agent training and inference are very fast, the former taking less than a minute. Despite the simplification inherent in this reformulated problem, our agent demonstrates performance that is at least on par with established methods. This assertion is validated by testing it across standard datasets associated with D4RL Gym-MuJoCo tasks. We further discuss the agent's ability to generalize by testing it on two extreme cases, how it learns to model the return distributions effectively even with highly skewed expert datasets, and how it exhibits robust performance in scenarios with sparse/delayed rewards.

Reinforcement Learning (RL) consists of designing agents that make intelligent decisions without human supervision. When used alongside function approximators such as Neural Networks (NNs), RL is capable of solving extremely complex problems. Deep Q-Learning, a RL algorithm that uses Deep NNs, achieved super-human performance in some specific tasks. Nonetheless, it is also possible to use Variational Quantum Circuits (VQCs) as function approximators in RL algorithms. This work empirically studies the performance and trainability of such VQC-based Deep Q-Learning models in classic control benchmark environments. More specifically, we research how data re-uploading affects both these metrics. We show that the magnitude and the variance of the gradients of these models remain substantial throughout training due to the moving targets of Deep Q-Learning. Moreover, we empirically show that increasing the number of qubits does not lead to an exponential vanishing behavior of the magnitude and variance of the gradients for a PQC approximating a 2-design, unlike what was expected due to the Barren Plateau Phenomenon. This hints at the possibility of VQCs being specially adequate for being used as function approximators in such a context.