Inverse reinforcement learning is the problem of inferring a reward function from an optimal policy. In this work, it is assumed that the reward is expressed as a reward machine whose transitions depend on atomic propositions associated with the state of a Markov Decision Process (MDP). Our goal is to identify the true reward machine using finite information. To this end, we first introduce the notion of a prefix tree policy which associates a distribution of actions to each state of the MDP and each attainable finite sequence of atomic propositions. Then, we characterize an equivalence class of reward machines that can be identified given the prefix tree policy. Finally, we propose a SAT-based algorithm that uses information extracted from the prefix tree policy to solve for a reward machine. It is proved that if the prefix tree policy is known up to a sufficient (but finite) depth, our algorithm recovers the exact reward machine up to the equivalence class. This sufficient depth is derived as a function of the number of MDP states and (an upper bound on) the number of states of the reward machine. Several examples are used to demonstrate the effectiveness of the approach.

Even highly capable large language models (LLMs) can produce biased or unsafe responses, and alignment techniques, such as RLHF, aimed at mitigating this issue, are expensive and prone to overfitting as they retrain the LLM. This paper introduces a novel inference-time alignment approach that ensures LLMs generate safe responses almost surely, i.e., with a probability approaching one. We achieve this by framing the safe generation of inference-time responses as a constrained Markov decision process within the LLM's latent space. Crucially, we augment a safety state that tracks the evolution of safety constraints and enables us to demonstrate formal safety guarantees upon solving the MDP in the latent space. Building on this foundation, we propose InferenceGuard, a practical implementation that safely aligns LLMs without modifying the model weights. Empirically, we demonstrate InferenceGuard effectively balances safety and task performance, outperforming existing inference-time alignment methods in generating safe and aligned responses.

Value Iteration is a widely used algorithm for solving Markov Decision Processes (MDPs). While previous studies have extensively analyzed its convergence properties, they primarily focus on convergence with respect to the infinity norm. In this work, we use absolute probability sequences to develop a new line of analysis and examine the algorithm's convergence in terms of the $L^2$ norm, offering a new perspective on its behavior and performance.

Flocking control is essential for multi-robot systems in diverse applications, yet achieving efficient flocking in congested environments poses challenges regarding computation burdens, performance optimality, and motion safety. This paper addresses these challenges through a multi-agent reinforcement learning (MARL) framework built on Gibbs Random Fields (GRFs). With GRFs, a multi-robot system is represented by a set of random variables conforming to a joint probability distribution, thus offering a fresh perspective on flocking reward design. A decentralized training and execution mechanism, which enhances the scalability of MARL concerning robot quantity, is realized using a GRF-based credit assignment method. An action attention module is introduced to implicitly anticipate the motion intentions of neighboring robots, consequently mitigating potential non-stationarity issues in MARL. The proposed framework enables learning an efficient distributed control policy for multi-robot systems in challenging environments with success rate around $99\%$, as demonstrated through thorough comparisons with state-of-the-art solutions in simulations and experiments. Ablation studies are also performed to validate the efficiency of different framework modules.

Predicting the future trajectories of surrounding traffic participants plays an essential role in automated driving. By anticipating future movements of nearby agents, such as vehicles and vulnerable road users, an automated vehicle (AV) can better plan maneuvers, reduce the risk of collisions, and ensure smoother interactions with other road users. Although existing approaches, e.g., [1-3], effectively predict the future movements of individual traffic participants, they limit an AV to a reactive planning strategy, assuming that the predictions of surrounding vehicles remain unaffected by the AV's planned actions. In highly interactive situations, this often leads to the freezing robot problem [4], where the AV, unable to engage in cooperative planning, simply stops to avoid potential collisions. For example, when it is unable to merge in dense traffic because the predictions of surrounding vehicles do not react to the AV's plan. One approach to resolving this is to condition the prediction on the AV's plan, often referred to as conditional inference [5].

In this paper, we study an online algorithm in a reproducing kernel Hilbert spaces (RKHS) based on a class of dependent processes, called the mixing process. For such a process, the degree of dependence is measured by various mixing coefficients. As a representative example, we analyze a strictly stationary Markov chain, where the dependence structure is characterized by the \(\beta-\) and \(\phi-\)mixing coefficients. For these dependent samples, we derive nearly optimal convergence rates. Our findings extend existing error bounds for i.i.d. observations, demonstrating that the i.i.d. case is a special instance of our framework. Moreover, we explicitly account for an additional factor introduced by the dependence structure in the Markov chain.

Multi-agent reinforcement learning typically employs a centralized training-decentralized execution (CTDE) framework to alleviate the non-stationarity in environment. However, the partial observability during execution may lead to cumulative gap errors gathered by agents, impairing the training of effective collaborative policies. To overcome this challenge, we introduce the Double Distillation Network (DDN), which incorporates two distillation modules aimed at enhancing robust coordination and facilitating the collaboration process under constrained information. The external distillation module uses a global guiding network and a local policy network, employing distillation to reconcile the gap between global training and local execution. In addition, the internal distillation module introduces intrinsic rewards, drawn from state information, to enhance the exploration capabilities of agents. Extensive experiments demonstrate that DDN significantly improves performance across multiple scenarios.

Large Language Models (LLM)-based systems, i.e. interconnected elements that include an LLM as a central component (e.g., conversational agents), are typically monolithic static architectures that rely on a single LLM for all user queries. However, they often require different preprocessing strategies, levels of reasoning, or knowledge. Generalist LLMs (e.g. GPT-4) trained on very large multi-topic corpora can perform well in a variety of tasks. They require significant financial, energy, and hardware resources that may not be justified for basic tasks. This implies potentially investing in unnecessary costs for a given query. To overcome this problem, a routing mechanism routes user queries to the most suitable components, such as smaller LLMs or experts in specific topics. This approach may improve response quality while minimising costs. Routing can be expanded to other components of the conversational agent architecture, such as the selection of optimal embedding strategies. This paper explores key considerations for integrating routing into LLM-based systems, focusing on resource management, cost definition, and strategy selection. Our main contributions include a formalisation of the problem, a novel taxonomy of existing approaches emphasising relevance and resource efficiency, and a comparative analysis of these strategies in relation to industry practices. Finally, we identify critical challenges and directions for future research.

The fields of MPC and RL consider two successful control techniques for Markov decision processes. Both approaches are derived from similar fundamental principles, and both are widely used in practical applications, including robotics, process control, energy systems, and autonomous driving. Despite their similarities, MPC and RL follow distinct paradigms that emerged from diverse communities and different requirements. Various technical discrepancies, particularly the role of an environment model as part of the algorithm, lead to methodologies with nearly complementary advantages. Due to their orthogonal benefits, research interest in combination methods has recently increased significantly, leading to a large and growing set of complex ideas leveraging MPC and RL. This work illuminates the differences, similarities, and fundamentals that allow for different combination algorithms and categorizes existing work accordingly. Particularly, we focus on the versatile actor-critic RL approach as a basis for our categorization and examine how the online optimization approach of MPC can be used to improve the overall closed-loop performance of a policy.

The main works [10, 11] attack the same realistic fraud detection Adversarial attacks pose a significant threat to data-driven engine called BankSealer [9]. In both works, the authors systems, and researchers have spent considerable resources rightfully consider domain-specific challenges generally absent studying them. Despite its economic relevance, this trend in other adversarial works, such as the intricate feature largely overlooked the issue of credit card fraud detection. To engineering process performed in fraud detection. However, address this gap, we propose a new threat model that demonstrates they operate under the assumption that fraudsters can access the limitations of existing attacks and highlights the the customers' transaction history. As the authors point out, necessity to investigate new approaches. We then design a this may be achieved through the introduction of malware into new adversarial attack for credit card fraud detection, employing the victim's devices. However, this considerably increases the reinforcement learning to bypass classifiers. This attack, difficulty of performing any attack, as fraudsters must first called FRAUD-RLA, is designed to maximize the attacker's compromise the customer's device and observe past transaction reward by optimizing the exploration-exploitation tradeoff history, which constitutes a significantly more complex and working with significantly less required knowledge than undertaking than stealing or cloning a card.