In retail (e.g., grocery stores, apparel shops, online retailers), inventory managers have to balance short-term risk (no items to sell) with long-term-risk (over ordering leading to product waste). This balancing task is made especially hard due to the lack of information about future customer purchases. In this paper, we study the problem of restocking a grocery store's inventory with perishable items over time, from a distributional point of view. The objective is to maximize sales while minimizing waste, with uncertainty about the actual consumption by costumers. This problem is of a high relevance today, given the growing demand for food and the impact of food waste on the environment, the economy, and purchasing power. We frame inventory restocking as a new reinforcement learning task that exhibits stochastic behavior conditioned on the agent's actions, making the environment partially observable. We make two main contributions. First, we introduce a new reinforcement learning environment, RetaiL, based on real grocery store data and expert knowledge. This environment is highly stochastic, and presents a unique challenge for reinforcement learning practitioners. We show that uncertainty about the future behavior of the environment is not handled well by classical supply chain algorithms, and that distributional approaches are a good way to account for the uncertainty. Second, we introduce GTDQN, a distributional reinforcement learning algorithm that learns a generalized Tukey Lambda distribution over the reward space. GTDQN provides a strong baseline for our environment. It outperforms other distributional reinforcement learning approaches in this partially observable setting, in both overall reward and reduction of generated waste.

Offline reinforcement learning (RL) have received rising interest due to its appealing data efficiency. The present study addresses behavior estimation, a task that lays the foundation of many offline RL algorithms. Behavior estimation aims at estimating the policy with which training data are generated. In particular, this work considers a scenario where the data are collected from multiple sources. In this case, neglecting data heterogeneity, existing approaches for behavior estimation suffers from behavior misspecification. To overcome this drawback, the present study proposes a latent variable model to infer a set of policies from data, which allows an agent to use as behavior policy the policy that best describes a particular trajectory. This model provides with a agent fine-grained characterization for multi-source data and helps it overcome behavior misspecification. This work also proposes a learning algorithm for this model and illustrates its practical usage via extending an existing offline RL algorithm. Lastly, with extensive evaluation this work confirms the existence of behavior misspecification and the efficacy of the proposed model.

Object goal navigation is an important problem in Embodied AI that involves guiding the agent to navigate to an instance of the object category in an unknown environment -- typically an indoor scene. Unfortunately, current state-of-the-art methods for this problem rely heavily on data-driven approaches, \eg, end-to-end reinforcement learning, imitation learning, and others. Moreover, such methods are typically costly to train and difficult to debug, leading to a lack of transferability and explainability. Inspired by recent successes in combining classical and learning methods, we present a modular and training-free solution, which embraces more classic approaches, to tackle the object goal navigation problem. Our method builds a structured scene representation based on the classic visual simultaneous localization and mapping (V-SLAM) framework. We then inject semantics into geometric-based frontier exploration to reason about promising areas to search for a goal object. Our structured scene representation comprises a 2D occupancy map, semantic point cloud, and spatial scene graph. Our method propagates semantics on the scene graphs based on language priors and scene statistics to introduce semantic knowledge to the geometric frontiers. With injected semantic priors, the agent can reason about the most promising frontier to explore. The proposed pipeline shows strong experimental performance for object goal navigation on the Gibson benchmark dataset, outperforming the previous state-of-the-art. We also perform comprehensive ablation studies to identify the current bottleneck in the object navigation task.

Recent research in offline reinforcement learning (RL) has demonstrated that return-conditioned supervised learning is a powerful paradigm for decision-making problems. While promising, return conditioning is limited to training data labeled with rewards and therefore faces challenges in learning from unsupervised data. In this work, we aim to utilize generalized future conditioning to enable efficient unsupervised pretraining from reward-free and sub-optimal offline data. We propose Pretrained Decision Transformer (PDT), a conceptually simple approach for unsupervised RL pretraining. PDT leverages future trajectory information as a privileged context to predict actions during training. The ability to make decisions based on both present and future factors enhances PDT's capability for generalization. Besides, this feature can be easily incorporated into a return-conditioned framework for online finetuning, by assigning return values to possible futures and sampling future embeddings based on their respective values. Empirically, PDT outperforms or performs on par with its supervised pretraining counterpart, especially when dealing with sub-optimal data. Further analysis reveals that PDT can extract diverse behaviors from offline data and controllably sample high-return behaviors by online finetuning. Code is available at here.

Offline reinforcement learning (RL) allows agents to learn effective, return-maximizing policies from a static dataset. Three major paradigms for offline RL are Q-Learning, Imitation Learning, and Sequence Modeling. A key open question is: which paradigm is preferred under what conditions? We study this question empirically by exploring the performance of representative algorithms -- Conservative Q-Learning (CQL), Behavior Cloning (BC), and Decision Transformer (DT) -- across the commonly used D4RL and Robomimic benchmarks. We design targeted experiments to understand their behavior concerning data suboptimality and task complexity. Our key findings are: (1) Sequence Modeling requires more data than Q-Learning to learn competitive policies but is more robust; (2) Sequence Modeling is a substantially better choice than both Q-Learning and Imitation Learning in sparse-reward and low-quality data settings; and (3) Sequence Modeling and Imitation Learning are preferable as task horizon increases, or when data is obtained from human demonstrators. Based on the overall strength of Sequence Modeling, we also investigate architectural choices and scaling trends for DT on Atari and D4RL and make design recommendations. We find that scaling the amount of data for DT by 5x gives a 2.5x average score improvement on Atari.

Voltage control is crucial to large-scale power system reliable operation, as timely reactive power support can help prevent widespread outages. However, there is currently no built in mechanism for power systems to ensure that the voltage control objective to maintain reliable operation will survive or sustain the uncertainty caused under adversary presence. Hence, this work introduces a Bayesian Reinforcement Learning (BRL) approach for power system control problems, with focus on sustained voltage control under uncertainty in a cyber-adversarial environment. This work proposes a data-driven BRL-based approach for automatic voltage control by formulating and solving a Partially-Observable Markov Decision Problem (POMDP), where the states are partially observable due to cyber intrusions. The techniques are evaluated on the WSCC and IEEE 14 bus systems. Additionally, BRL techniques assist in automatically finding a threshold for exploration and exploitation in various RL techniques.

We analyse quantile temporal-difference learning (QTD), a distributional reinforcement learning algorithm that has proven to be a key component in several successful large-scale applications of reinforcement learning. Despite these empirical successes, a theoretical understanding of QTD has proven elusive until now. Unlike classical TD learning, which can be analysed with standard stochastic approximation tools, QTD updates do not approximate contraction mappings, are highly non-linear, and may have multiple fixed points. The core result of this paper is a proof of convergence to the fixed points of a related family of dynamic programming procedures with probability 1, putting QTD on firm theoretical footing. The proof establishes connections between QTD and non-linear differential inclusions through stochastic approximation theory and non-smooth analysis.

We propose Convex Constraint Learning for Reinforcement Learning (CoCoRL), a novel approach for inferring shared constraints in a Constrained Markov Decision Process (CMDP) from a set of safe demonstrations with possibly different reward functions. While previous work is limited to demonstrations with known rewards or fully known environment dynamics, CoCoRL can learn constraints from demonstrations with different unknown rewards without knowledge of the environment dynamics. CoCoRL constructs a convex safe set based on demonstrations, which provably guarantees safety even for potentially sub-optimal (but safe) demonstrations. For near-optimal demonstrations, CoCoRL converges to the true safe set with no policy regret. We evaluate CoCoRL in tabular environments and a continuous driving simulation with multiple constraints. CoCoRL learns constraints that lead to safe driving behavior and that can be transferred to different tasks and environments. In contrast, alternative methods based on Inverse Reinforcement Learning (IRL) often exhibit poor performance and learn unsafe policies.

Large transformer models powered by diverse data and model scale have dominated natural language modeling and computer vision and pushed the frontier of multiple AI areas. In reinforcement learning (RL), despite many efforts into transformer-based policies, a key limitation, however, is that current transformer-based policies cannot learn by directly combining information from multiple sub-optimal trials. In this work, we address this issue using recently proposed chain of hindsight to relabel experience, where we train a transformer on a sequence of trajectory experience ascending sorted according to their total rewards. Our method consists of relabelling target return of each trajectory to the maximum total reward among in sequence of trajectories and training an autoregressive model to predict actions conditioning on past states, actions, rewards, target returns, and task completion tokens, the resulting model, Agentic Transformer (AT), can learn to improve upon itself both at training and test time. As we show on D4RL and ExoRL benchmarks, to the best our knowledge, this is the first time that a simple transformer-based model performs competitively with both temporal-difference and imitation-learning-based approaches, even from sub-optimal data. Our Agentic Transformer also shows a promising scaling trend that bigger models consistently improve results.

We provide a new non-asymptotic analysis of distributed temporal difference learning with linear function approximation. Our approach relies on ``one-shot averaging,'' where $N$ agents run identical local copies of the TD(0) method and average the outcomes only once at the very end. We demonstrate a version of the linear time speedup phenomenon, where the convergence time of the distributed process is a factor of $N$ faster than the convergence time of TD(0). This is the first result proving benefits from parallelism for temporal difference methods.