Deep reinforcement learning (RL) has shown remarkable success in specific offline decision-making scenarios, yet its theoretical guarantees are still under development. Existing works on offline RL theory primarily emphasize a few trivial settings, such as linear MDP or general function approximation with strong assumptions and independent data, which lack guidance for practical use. The coupling of deep learning and Bellman residuals makes this problem challenging, in addition to the difficulty of data dependence. In this paper, we establish a non-asymptotic estimation error of pessimistic offline RL using general neural network approximation with $\mathcal{C}$-mixing data regarding the structure of networks, the dimension of datasets, and the concentrability of data coverage, under mild assumptions. Our result shows that the estimation error consists of two parts: the first converges to zero at a desired rate on the sample size with partially controllable concentrability, and the second becomes negligible if the residual constraint is tight. This result demonstrates the explicit efficiency of deep adversarial offline RL frameworks. We utilize the empirical process tool for $\mathcal{C}$-mixing sequences and the neural network approximation theory for the H\"{o}lder class to achieve this. We also develop methods to bound the Bellman estimation error caused by function approximation with empirical Bellman constraint perturbations. Additionally, we present a result that lessens the curse of dimensionality using data with low intrinsic dimensionality and function classes with low complexity. Our estimation provides valuable insights into the development of deep offline RL and guidance for algorithm model design.

Long-tailed (LT) classification is an unavoidable and challenging problem in the real world. Most existing long-tailed classification methods focus only on solving the class-wise imbalance while ignoring the attribute-wise imbalance. The deviation of a classification model is caused by both class-wise and attribute-wise imbalance. Due to the fact that attributes are implicit in most datasets and the combination of attributes is complex, attribute-wise imbalance is more difficult to handle. For this purpose, we proposed a novel long-tailed classification framework, aiming to build a multi-granularity classification model by means of invariant feature learning. This method first unsupervisedly constructs Coarse-Grained forest (CLF) to better characterize the distribution of attributes within a class. Depending on the distribution of attributes, one can customize suitable sampling strategies to construct different imbalanced datasets. We then introduce multi-center loss (MCL) that aims to gradually eliminate confusing attributes during feature learning process. The proposed framework does not necessarily couple to a specific LT classification model structure and can be integrated with any existing LT method as an independent component. Extensive experiments show that our approach achieves state-of-the-art performance on both existing benchmarks ImageNet-GLT and MSCOCO-GLT and can improve the performance of existing LT methods. Our codes are available on GitHub: \url{https://github.com/jinyery/cognisance}

Efficient traffic signal control is critical for reducing traffic congestion and improving overall transportation efficiency. The dynamic nature of traffic flow has prompted researchers to explore Reinforcement Learning (RL) for traffic signal control (TSC). Compared with traditional methods, RL-based solutions have shown preferable performance. However, the application of RL-based traffic signal controllers in the real world is limited by the low sample efficiency and high computational requirements of these solutions. In this work, we propose DTLight, a simple yet powerful lightweight Decision Transformer-based TSC method that can learn policy from easily accessible offline datasets. DTLight novelly leverages knowledge distillation to learn a lightweight controller from a well-trained larger teacher model to reduce implementation computation. Additionally, it integrates adapter modules to mitigate the expenses associated with fine-tuning, which makes DTLight practical for online adaptation with minimal computation and only a few fine-tuning steps during real deployment. Moreover, DTLight is further enhanced to be more applicable to real-world TSC problems. Extensive experiments on synthetic and real-world scenarios show that DTLight pre-trained purely on offline datasets can outperform state-of-the-art online RL-based methods in most scenarios. Experiment results also show that online fine-tuning further improves the performance of DTLight by up to 42.6% over the best online RL baseline methods. In this work, we also introduce Datasets specifically designed for TSC with offline RL (referred to as DTRL). Our datasets and code are publicly available.

The use of learning-based methods for optimizing cellular radio access networks (RAN) has received increasing attention in recent years. This coincides with a rapid increase in the number of cell sites worldwide, driven largely by dramatic growth in cellular network traffic. Training and maintaining learned models that work well across a large number of cell sites has thus become a pertinent problem. This paper proposes a scalable framework for constructing a reinforcement learning policy bank that can perform RAN optimization across a large number of cell sites with varying traffic patterns. Central to our framework is a novel application of anomaly detection techniques to assess the compatibility between sites (tasks) and the policy bank. This allows our framework to intelligently identify when a policy can be reused for a task, and when a new policy needs to be trained and added to the policy bank. Our results show that our approach to compatibility assessment leads to an efficient use of computational resources, by allowing us to construct a performant policy bank without exhaustively training on all tasks, which makes it applicable under real-world constraints.

In this paper, we study the \textit{graph condensation} problem by compressing the large, complex graph into a concise, synthetic representation that preserves the most essential and discriminative information of structure and features. We seminally propose the concept of Shock Absorber (a type of perturbation) that enhances the robustness and stability of the original graphs against changes in an adversarial training fashion. Concretely, (I) we forcibly match the gradients between pre-selected graph neural networks (GNNs) trained on a synthetic, simplified graph and the original training graph at regularly spaced intervals. (II) Before each update synthetic graph point, a Shock Absorber serves as a gradient attacker to maximize the distance between the synthetic dataset and the original graph by selectively perturbing the parts that are underrepresented or insufficiently informative. We iteratively repeat the above two processes (I and II) in an adversarial training fashion to maintain the highly-informative context without losing correlation with the original dataset. More importantly, our shock absorber and the synthesized graph parallelly share the backward process in a free training manner. Compared to the original adversarial training, it introduces almost no additional time overhead. We validate our framework across 8 datasets (3 graph and 5 node classification datasets) and achieve prominent results: for example, on Cora, Citeseer and Ogbn-Arxiv, we can gain nearly 1.13% to 5.03% improvements compare with SOTA models. Moreover, our algorithm adds only about 0.2% to 2.2% additional time overhead over Flicker, Citeseer and Ogbn-Arxiv. Compared to the general adversarial training, our approach improves time efficiency by nearly 4-fold.

This paper aims to propose an online two-stage thermal history prediction method, which could be integrated into a metal AM process for performance control. Based on the similarity of temperature curves (curve segments of a temperature profile of one point) between any two successive layers, the first stage of the proposed method designs a layer-to-layer prediction model to estimate the temperature curves of the yet-to-print layer from measured temperatures of certain points on the previously printed layer. With measured/predicted temperature profiles of several points on the same layer, the second stage proposes a reduced order model (ROM) (intra-layer prediction model) to decompose and construct the temperature profiles of all points on the same layer, which could be used to build the temperature field of the entire layer. The training of ROM is performed with an extreme learning machine (ELM) for computational efficiency. Fifteen wire arc AM experiments and nine simulations are designed for thin walls with a fixed length and unidirectional printing of each layer. The test results indicate that the proposed prediction method could construct the thermal history of a yet-to-print layer within 0.1 seconds on a low-cost desktop computer. Meanwhile, the method has acceptable generalization capability in most cases from lower layers to higher layers in the same simulation, as well as from one simulation to a new simulation on different AM process parameters. More importantly, after fine-tuning the proposed method with limited experimental data, the relative errors of all predicted temperature profiles on a new experiment are smaller than 0.09, which demonstrates the applicability and generalization of the proposed two-stage thermal history prediction method in online applications for metal AM.

Knowledge graph embedding (KGE) that maps entities and relations into vector representations is essential for downstream applications. Conventional KGE methods require high-dimensional representations to learn the complex structure of knowledge graph, but lead to oversized model parameters. Recent advances reduce parameters by low-dimensional entity representations, while developing techniques (e.g., knowledge distillation or reinvented representation forms) to compensate for reduced dimension. However, such operations introduce complicated computations and model designs that may not benefit large knowledge graphs. To seek a simple strategy to improve the parameter efficiency of conventional KGE models, we take inspiration from that deeper neural networks require exponentially fewer parameters to achieve expressiveness comparable to wider networks for compositional structures. We view all entity representations as a single-layer embedding network, and conventional KGE methods that adopt high-dimensional entity representations equal widening the embedding network to gain expressiveness. To achieve parameter efficiency, we instead propose a deeper embedding network for entity representations, i.e., a narrow entity embedding layer plus a multi-layer dimension lifting network (LiftNet). Experiments on three public datasets show that by integrating LiftNet, four conventional KGE methods with 16-dimensional representations achieve comparable link prediction accuracy as original models that adopt 512-dimensional representations, saving 68.4% to 96.9% parameters.

To better match drivers to riders in our ridesharing application, we revised Lyft's core matching algorithm. We use a novel online reinforcement learning approach that estimates the future earnings of drivers in real time and use this information to find more efficient matches. This change was the first documented implementation of a ridesharing matching algorithm that can learn and improve in real time. We evaluated the new approach during weeks of switchback experimentation in most Lyft markets, and estimated how it benefited drivers, riders, and the platform. In particular, it enabled our drivers to serve millions of additional riders each year, leading to more than $30 million per year in incremental revenue. Lyft rolled out the algorithm globally in 2021.

Deep learning has been extensively used in wireless communication problems, including channel estimation. Although several data-driven approaches exist, a fair and realistic comparison between them is difficult due to inconsistencies in the experimental conditions and the lack of a standardized experimental design. In addition, the performance of data-driven approaches is often compared based on empirical analysis. The lack of reproducibility and availability of standardized evaluation tools (e.g., datasets, codebases) hinder the development and progress of data-driven methods for channel estimation and wireless communication in general. In this work, we introduce an initiative to build benchmarks that unify several data-driven OFDM channel estimation approaches. Specifically, we present CeBed (a testbed for channel estimation) including different datasets covering various systems models and propagation conditions along with the implementation of ten deep and traditional baselines. This benchmark considers different practical aspects such as the robustness of the data-driven models, the number and the arrangement of pilots, and the number of receive antennas. This work offers a comprehensive and unified framework to help researchers evaluate and design data-driven channel estimation algorithms.

Instruction tuning (IT) achieves impressive zero-shot generalization results by training large language models (LLMs) on a massive amount of diverse tasks with instructions. However, how to select new tasks to improve the performance and generalizability of IT models remains an open question. Training on all existing tasks is impractical due to prohibiting computation requirements, and randomly selecting tasks can lead to suboptimal performance. In this work, we propose active instruction tuning based on prompt uncertainty, a novel framework to identify informative tasks, and then actively tune the models on the selected tasks. We represent the informativeness of new tasks with the disagreement of the current model outputs over perturbed prompts. Our experiments on NIV2 and Self-Instruct datasets demonstrate that our method consistently outperforms other baseline strategies for task selection, achieving better out-of-distribution generalization with fewer training tasks. Additionally, we introduce a task map that categorizes and diagnoses tasks based on prompt uncertainty and prediction probability. We discover that training on ambiguous (prompt-uncertain) tasks improves generalization while training on difficult (prompt-certain and low-probability) tasks offers no benefit, underscoring the importance of task selection for instruction tuning.