For continuing tasks, average cost Markov decision processes have well-documented value and can be solved using efficient algorithms. However, it explicitly assumes that the agent is risk-neutral. In this work, we extend risk-neutral algorithms to accommodate the more general class of dynamic risk measures. Specifically, we propose a relative value iteration (RVI) algorithm for planning and design two model-free Q-learning algorithms, namely a generic algorithm based on the multi-level Monte Carlo method, and an off-policy algorithm dedicated to utility-base shortfall risk measures. Both the RVI and MLMC-based Q-learning algorithms are proven to converge to optimality. Numerical experiments validate our analysis, confirms empirically the convergence of the off-policy algorithm, and demonstrate that our approach enables the identification of policies that are finely tuned to the intricate risk-awareness of the agent that they serve.

Applying imitation learning (IL) is challenging to nonprehensile manipulation tasks of invisible objects with partial observations, such as excavating buried rocks. The demonstrator must make such complex action decisions as exploring to find the object and task-oriented actions to complete the task while estimating its hidden state, perhaps causing inconsistent action demonstration and high cognitive load problems. For these problems, work in human cognitive science suggests that promoting the use of pre-designed, simple exploration rules for the demonstrator may alleviate the problems of action inconsistency and high cognitive load. Therefore, when performing imitation learning from demonstrations using such exploration rules, it is important to accurately imitate not only the demonstrator's task-oriented behavior but also his/her mode-switching behavior (exploratory or task-oriented behavior) under partial observation. Based on the above considerations, this paper proposes a novel imitation learning framework called Belief Exploration-Action Cloning (BEAC), which has a switching policy structure between a pre-designed exploration policy and a task-oriented action policy trained on the estimated belief states based on past history. In simulation and real robot experiments, we confirmed that our proposed method achieved the best task performance, higher mode and action prediction accuracies, while reducing the cognitive load in the demonstration indicated by a user study.

Large language models (LLMs) have recently transformed from text-based assistants to autonomous agents capable of planning, reasoning, and iteratively improving their actions. While numerical reward signals and verifiers can effectively rank candidate actions, they often provide limited contextual guidance. In contrast, natural language feedback better aligns with the generative capabilities of LLMs, providing richer and more actionable suggestions. However, parsing and implementing this feedback effectively can be challenging for LLM-based agents. In this work, we introduce Critique-Guided Improvement (CGI), a novel two-player framework, comprising an actor model that explores an environment and a critic model that generates detailed nature language feedback. By training the critic to produce fine-grained assessments and actionable revisions, and the actor to utilize these critiques, our approach promotes more robust exploration of alternative strategies while avoiding local optima. Experiments in three interactive environments show that CGI outperforms existing baselines by a substantial margin. Notably, even a small critic model surpasses GPT-4 in feedback quality. The resulting actor achieves state-of-the-art performance, demonstrating the power of explicit iterative guidance to enhance decision-making in LLM-based agents.

We propose Diffusion-Informed Model Predictive Control (D-I MPC), a generic framework for uncertainty-aware prediction and decision-making in partially observable stochastic systems by integrating diffusion-based time series forecasting models in Model Predictive Control algorithms. In our approach, a diffusion-based time series forecasting model is used to probabilistically estimate the evolution of the system's stochastic components. These forecasts are then incorporated into MPC algorithms to estimate future trajectories and optimize action selection under the uncertainty of the future. We evaluate the framework on the task of energy arbitrage, where a Battery Energy Storage System participates in the day-ahead electricity market of the New York state. Experimental results indicate that our model-based approach with a diffusion-based forecaster significantly outperforms both implementations with classical forecasting methods and model-free reinforcement learning baselines.

Multi-Agent Deep Reinforcement Learning (MADRL) has emerged as a powerful tool for optimizing decentralized decision-making systems in complex settings, such as Dynamic Spectrum Access (DSA). However, deploying deep learning models on resource-constrained edge devices remains challenging due to their high computational cost. To address this challenge, in this paper, we present a novel sparse recurrent MARL framework integrating gradual neural network pruning into the independent actor global critic paradigm. Additionally, we introduce a harmonic annealing sparsity scheduler, which achieves comparable, and in certain cases superior, performance to standard linear and polynomial pruning schedulers at large sparsities. Our experimental investigation demonstrates that the proposed DSA framework can discover superior policies, under diverse training conditions, outperforming conventional DSA, MADRL baselines, and state-of-the-art pruning techniques.

This study presents a robust optimization algorithm for automated highway merge. The merging scenario is one of the challenging scenes in automated driving, because it requires adjusting ego vehicle's speed to match other vehicles before reaching the end point. Then, we model the speed planning problem as a deterministic Markov decision process. The proposed scheme is able to compute each state value of the process and reliably derive the optimal sequence of actions. In our approach, we adopt jerk as the action of the process to prevent a sudden change of acceleration. However, since this expands the state space, we also consider ways to achieve a real-time operation. We compared our scheme with a simple algorithm with the Intelligent Driver Model. We not only evaluated the scheme in a simulation environment but also conduct a real world testing.

Large language models (LLMs) like GPT-4 and LLaMA-3 utilize the powerful in-context learning (ICL) capability of Transformer architecture to learn on the fly from limited examples. While ICL underpins many LLM applications, its full potential remains hindered by a limited understanding of its generalization boundaries and vulnerabilities. We present a systematic investigation of transformers' generalization capability with ICL relative to training data coverage by defining a task-centric framework along three dimensions: inter-problem, intra-problem, and intra-task generalization. Through extensive simulation and real-world experiments, encompassing tasks such as function fitting, API calling, and translation, we find that transformers lack inter-problem generalization with ICL, but excel in intra-task and intra-problem generalization. When the training data includes a greater variety of mixed tasks, it significantly enhances the generalization ability of ICL on unseen tasks and even on known simple tasks. This guides us in designing training data to maximize the diversity of tasks covered and to combine different tasks whenever possible, rather than solely focusing on the target task for testing.

Partially observable environments present a considerable computational challenge in reinforcement learning due to the need to consider long histories. Learning with a finite window of observations quickly becomes intractable as the window length grows. In this work, we introduce memory traces. Inspired by eligibility traces, these are compact representations of the history of observations in the form of exponential moving averages. We prove sample complexity bounds for the problem of offline on-policy evaluation that quantify the value errors achieved with memory traces for the class of Lipschitz continuous value estimates. We establish a close connection to the window approach, and demonstrate that, in certain environments, learning with memory traces is significantly more sample efficient. Finally, we underline the effectiveness of memory traces empirically in online reinforcement learning experiments for both value prediction and control.

Foundation models (FMs) have achieved remarkable success across various domains, yet their adoption in healthcare remains limited. While significant advances have been made in medical imaging, genetic biomarkers, and time series from electronic health records, the potential of FMs for patient behavior monitoring through wearable devices remains underexplored. These datasets are inherently heterogeneous, multisource, and often exhibit high rates of missing data, posing unique challenges. This paper introduces a novel FM based on a modified vector quantized variational autoencoder (VQ-VAE), specifically designed to process real-world data from wearable devices. We demonstrate that our pretrained FM, trained on a broad cohort of psychiatric patients, performs downstream tasks via its latent representation without fine-tuning on a held-out cohort of suicidal patients. To illustrate this, we develop a probabilistic change-point detection algorithm for suicide detection and demonstrate the FM's effectiveness in predicting emotional states. Our results show that the discrete latent structure of the VQ-VAE outperforms a state-of-the-art Informer architecture in unsupervised suicide detection, while matching its performance in supervised emotion prediction when the latent dimensionality is increased, though at the cost of reduced unsupervised accuracy. This trade-off highlights the need for future FMs to integrate hybrid discrete-continuous structures for balanced performance across tasks.

Robotics Reinforcement Learning (RL) often relies on carefully engineered auxiliary rewards to supplement sparse primary learning objectives to compensate for the lack of large-scale, real-world, trial-and-error data. While these auxiliary rewards accelerate learning, they require significant engineering effort, may introduce human biases, and cannot adapt to the robot's evolving capabilities during training. In this paper, we introduce Reward Training Wheels (RTW), a teacher-student framework that automates auxiliary reward adaptation for robotics RL. To be specific, the RTW teacher dynamically adjusts auxiliary reward weights based on the student's evolving capabilities to determine which auxiliary reward aspects require more or less emphasis to improve the primary objective. We demonstrate RTW on two challenging robot tasks: navigation in highly constrained spaces and off-road vehicle mobility on vertically challenging terrain. In simulation, RTW outperforms expert-designed rewards by 2.35% in navigation success rate and improves off-road mobility performance by 122.62%, while achieving 35% and 3X faster training efficiency, respectively. Physical robot experiments further validate RTW's effectiveness, achieving a perfect success rate (5/5 trials vs. 2/5 for expert-designed rewards) and improving vehicle stability with up to 47.4% reduction in orientation angles.