In real-world scenarios, it is desirable for embodied agents to have the ability to leverage human language to gain explicit or implicit knowledge for learning tasks. Despite recent progress, most previous approaches adopt simple low-level instructions as language inputs, which may not reflect natural human communication. It's not clear how to incorporate rich language use to facilitate task learning. To address this question, this paper studies different types of language inputs in facilitating reinforcement learning (RL) embodied agents. More specifically, we examine how different levels of language informativeness (i.e., feedback on past behaviors and future guidance) and diversity (i.e., variation of language expressions) impact agent learning and inference. Our empirical results based on four RL benchmarks demonstrate that agents trained with diverse and informative language feedback can achieve enhanced generalization and fast adaptation to new tasks. These findings highlight the pivotal role of language use in teaching embodied agents new tasks in an open world. Project website: https://github.com/sled-group/Teachable_RL

Mixed precision quantization has become an important technique for enabling the execution of deep neural networks (DNNs) on limited resource computing platforms. Traditional quantization methods have primarily concentrated on maintaining neural network accuracy, either ignoring the impact of quantization on the robustness of the network, or using only empirical techniques for improving robustness. In contrast, techniques for robustness certification, which can provide strong guarantees about the robustness of DNNs have not been used during quantization due to their high computation cost. This paper introduces ARQ, an innovative mixed-precision quantization method that not only preserves the clean accuracy of the smoothed classifiers but also maintains their certified robustness. ARQ uses reinforcement learning to find accurate and robust DNN quantization, while efficiently leveraging randomized smoothing, a popular class of statistical DNN verification algorithms, to guide the search process. We compare ARQ with multiple state-of-the-art quantization techniques on several DNN architectures commonly used in quantization studies: ResNet-20 on CIFAR-10, ResNet-50 on ImageNet, and MobileNetV2 on ImageNet. We demonstrate that ARQ consistently performs better than these baselines across all the benchmarks and the input perturbation levels. In many cases, the performance of ARQ quantized networks can reach that of the original DNN with floating-point weights, but with only 1.5% instructions.

Path planning in high-dimensional spaces poses significant challenges, particularly in achieving both time efficiency and a fair success rate. To address these issues, we introduce a novel path-planning algorithm, Zonal RL-RRT, that leverages kd-tree partitioning to segment the map into zones while addressing zone connectivity, ensuring seamless transitions between zones. By breaking down the complex environment into multiple zones and using Q-learning as the high-level decision-maker, our algorithm achieves a 3x improvement in time efficiency compared to basic sampling methods such as RRT and RRT* in forest-like maps. Our approach outperforms heuristic-guided methods like BIT* and Informed RRT* by 1.5x in terms of runtime while maintaining robust and reliable success rates across 2D to 6D environments. Compared to learning-based methods like NeuralRRT* and MPNetSMP, as well as the heuristic RRT*J, our algorithm demonstrates, on average, 1.5x better performance in the same environments. We also evaluate the effectiveness of our approach through simulations of the UR10e arm manipulator in the MuJoCo environment. A key observation of our approach lies in its use of zone partitioning and Reinforcement Learning (RL) for adaptive high-level planning allowing the algorithm to accommodate flexible policies across diverse environments, making it a versatile tool for advanced path planning.

The cooperative driving technology of Connected and Autonomous Vehicles (CAVs) is crucial for improving the efficiency and safety of transportation systems. Learning-based methods, such as Multi-Agent Reinforcement Learning (MARL), have demonstrated strong capabilities in cooperative decision-making tasks. However, existing MARL approaches still face challenges in terms of learning efficiency and performance. In recent years, Large Language Models (LLMs) have rapidly advanced and shown remarkable abilities in various sequential decision-making tasks. To enhance the learning capabilities of cooperative agents while ensuring decision-making efficiency and cost-effectiveness, we propose LDPD, a language-driven policy distillation method for guiding MARL exploration. In this framework, a teacher agent based on LLM trains smaller student agents to achieve cooperative decision-making through its own decision-making demonstrations. The teacher agent enhances the observation information of CAVs and utilizes LLMs to perform complex cooperative decision-making reasoning, which also leverages carefully designed decision-making tools to achieve expert-level decisions, providing high-quality teaching experiences. The student agent then refines the teacher's prior knowledge into its own model through gradient policy updates. The experiments demonstrate that the students can rapidly improve their capabilities with minimal guidance from the teacher and eventually surpass the teacher's performance. Extensive experiments show that our approach demonstrates better performance and learning efficiency compared to baseline methods.

In Markov decision processes (MDPs), quantile risk measures such as Value-at-Risk are a standard metric for modeling RL agents' preferences for certain outcomes. This paper proposes a new Q-learning algorithm for quantile optimization in MDPs with strong convergence and performance guarantees. The algorithm leverages a new, simple dynamic program (DP) decomposition for quantile MDPs. Compared with prior work, our DP decomposition requires neither known transition probabilities nor solving complex saddle point equations and serves as a suitable foundation for other model-free RL algorithms. Our numerical results in tabular domains show that our Q-learning algorithm converges to its DP variant and outperforms earlier algorithms.

In this paper we propose a formal, model-agnostic meta-learning framework for safe reinforcement learning. Our framework is inspired by how parents safeguard their children across a progression of increasingly riskier tasks, imparting a sense of safety that is carried over from task to task. We model this as a meta-learning process where each task is synchronized with a safeguard that monitors safety and provides a reward signal to the agent. The safeguard is implemented as a finite-state machine based on a safety specification; the reward signal is formally shaped around this specification. The safety specification and its corresponding safeguard can be arbitrarily complex and non-Markovian, which adds flexibility to the training process and explainability to the learned policy. The design of the safeguard is manual but it is high-level and model-agnostic, which gives rise to an end-to-end safe learning approach with wide applicability, from pixel-level game control to language model fine-tuning. Starting from a given set of safety specifications (tasks), we train a model such that it can adapt to new specifications using only a small number of training samples. This is made possible by our method for efficiently transferring safety bias between tasks, which effectively minimizes the number of safety violations. We evaluate our framework in a Minecraft-inspired Gridworld, a VizDoom game environment, and an LLM fine-tuning application. Agents trained with our approach achieve near-minimal safety violations, while baselines are shown to underperform.

Achieving the no-regret property for Reinforcement Learning (RL) problems in continuous state and action-space environments is one of the major open problems in the field. Existing solutions either work under very specific assumptions or achieve bounds that are vacuous in some regimes. Furthermore, many structural assumptions are known to suffer from a provably unavoidable exponential dependence on the time horizon $H$ in the regret, which makes any possible solution unfeasible in practice. In this paper, we identify local linearity as the feature that makes Markov Decision Processes (MDPs) both learnable (sublinear regret) and feasible (regret that is polynomial in $H$). We define a novel MDP representation class, namely Locally Linearizable MDPs, generalizing other representation classes like Linear MDPs and MDPS with low inherent Belmman error. Then, i) we introduce Cinderella, a no-regret algorithm for this general representation class, and ii) we show that all known learnable and feasible MDP families are representable in this class. We first show that all known feasible MDPs belong to a family that we call Mildly Smooth MDPs. Then, we show how any mildly smooth MDP can be represented as a Locally Linearizable MDP by an appropriate choice of representation. This way, Cinderella is shown to achieve state-of-the-art regret bounds for all previously known (and some new) continuous MDPs for which RL is learnable and feasible.

Hindsight experience replay (HER) is well-known to accelerate goal-based reinforcement learning (RL). While HER is generally applied to off-policy RL algorithms, we previously showed that HER can also accelerate on-policy algorithms, such as proximal policy optimization (PPO), for goal-based Predator-Prey environments. Here, we show that we can improve the previous PPO-HER algorithm by selectively applying HER in a principled manner.

The reasoning abilities of large language models (LLMs) have improved with chain-of-thought (CoT) prompting, allowing models to solve complex tasks in a stepwise manner. However, training CoT capabilities requires detailed reasoning data, which is often scarce. The self-taught reasoner (STaR) framework addresses this by using reinforcement learning to automatically generate reasoning steps, reducing reliance on human-labeled data. Although STaR and its variants have demonstrated empirical success, a theoretical foundation explaining these improvements is lacking. This work provides a theoretical framework for understanding the effectiveness of reinforcement learning on CoT reasoning and STaR. Our contributions are: (1) an analysis of policy improvement, showing why LLM reasoning improves iteratively with STaR; (2) conditions for convergence to an optimal reasoning policy; (3) an examination of STaR's robustness, explaining how it can improve reasoning even when incorporating occasional incorrect steps; and (4) criteria for the quality of pre-trained models necessary to initiate effective reasoning improvement. This framework aims to bridge empirical findings with theoretical insights, advancing reinforcement learning approaches for reasoning in LLMs.

This study investigates a counterintuitive phenomenon in adversarial machine learning: the potential for noise-based defenses to inadvertently aid evasion attacks in certain scenarios. While randomness is often employed as a defensive strategy against adversarial examples, our research reveals that this approach can sometimes backfire, particularly when facing adaptive attackers using reinforcement learning (RL). Our findings show that in specific cases, especially with visually noisy classes, the introduction of noise in the classifier's confidence values can be exploited by the RL attacker, leading to a significant increase in evasion success rates. In some instances, the noise-based defense scenario outperformed other strategies by up to 20\% on a subset of classes. However, this effect was not consistent across all classifiers tested, highlighting the complexity of the interaction between noise-based defenses and different models. These results suggest that in some cases, noise-based defenses can inadvertently create an adversarial training loop beneficial to the RL attacker. Our study emphasizes the need for a more nuanced approach to defensive strategies in adversarial machine learning, particularly in safety-critical applications. It challenges the assumption that randomness universally enhances defense against evasion attacks and highlights the importance of considering adaptive, RL-based attackers when designing robust defense mechanisms.