Recent advances in learning decision-making policies can largely be attributed to training expressive policy models, largely via imitation learning. While imitation learning discards non-expert data, reinforcement learning (RL) can still learn from suboptimal data. However, instantiating RL training of a new policy class often presents a different challenge: most deep RL machinery is co-developed with assumptions on the policy class and backbone, resulting in poor performance when the policy class changes. For instance, SAC utilizes a low-variance reparameterization policy gradient for Gaussian policies, but this is unstable for diffusion policies and intractable for autoregressive categorical policies. To address this issue, we develop an offline RL and online fine-tuning approach called policy-agnostic RL (PA-RL) that can effectively train multiple policy classes, with varying architectures and sizes. We build off the basic idea that a universal supervised learning loss can replace the policy improvement step in RL, as long as it is applied on "optimized" actions. To obtain these optimized actions, we first sample multiple actions from a base policy, and run global optimization (i.e., re-ranking multiple action samples using the Q-function) and local optimization (i.e., running gradient steps on an action sample) to maximize the critic on these candidates. PA-RL enables fine-tuning diffusion and transformer policies with either autoregressive tokens or continuous action outputs, at different sizes, entirely via actor-critic RL. Moreover, PA-RL improves the performance and sample-efficiency by up to 2 times compared to existing offline RL and online fine-tuning methods. We show the first result that successfully fine-tunes OpenVLA, a 7B generalist robot policy, autonomously with Cal-QL, an online RL fine-tuning algorithm, improving from 40% to 70% in the real world in 40 minutes.

Trilevel learning (TLL) found diverse applications in numerous machine learning applications, ranging from robust hyperparameter optimization to domain adaptation. However, existing researches primarily focus on scenarios where TLL can be addressed with first order information available at each level, which is inadequate in many situations involving zeroth order constraints, such as when black-box models are employed. Moreover, in trilevel learning, data may be distributed across various nodes, necessitating strategies to address TLL problems without centralizing data on servers to uphold data privacy. To this end, an effective distributed trilevel zeroth order learning framework DTZO is proposed in this work to address the TLL problems with level-wise zeroth order constraints in a distributed manner. The proposed DTZO is versatile and can be adapted to a wide range of (grey-box) TLL problems with partial zeroth order constraints. In DTZO, the cascaded polynomial approximation can be constructed without relying on gradients or sub-gradients, leveraging a novel cut, i.e., zeroth order cut. Furthermore, we theoretically carry out the non-asymptotic convergence rate analysis for the proposed DTZO in achieving the $\epsilon$-stationary point. Extensive experiments have been conducted to demonstrate and validate the superior performance of the proposed DTZO, e.g., it approximately achieves up to a 40$\%$ improvement in performance.

Due to the substantial rise in online retail and e-commerce in recent years, the demand for efficient and fast solutions to Vehicle Routing Problems (VRP) has become critical. To manage the increasing demand, companies have adopted the strategy of adding more depots. However, the presence of multiple depots introduces additional complexities, making existing VRP solutions suboptimal for addressing the Multi-depot Vehicle Routing Problem (MDVRP). Traditional methods for solving the MDVRP often require significant computation time, making them unsuitable for large-scale instances. Additionally, existing learning-based solutions for the MDVRP struggle with generalizability and fail to deliver high-quality results for scenarios involving a large number of customers. In this paper, we propose a novel solution for MDVRP. Our approach employs an attention mechanism, featuring a decoder with two key layers: one layer to consider the states of all vehicles and learn to select the most suitable vehicle based on the proximity of unassigned customers, and another layer to focus on assigning a customer to the selected vehicle. This approach delivers high-quality solutions for large-scale MDVRP instances and demonstrates remarkable generalizability across varying numbers of customers and depots. Its adaptability and performance make it a practical and deployable solution for real-world logistics challenges.

The rapid proliferation of Internet-of-things (IoT) as well as mobile devices such as Electric Vehicles (EVs), has led to unpredictable load at the grid. The demand to supply ratio is particularly exacerbated at a few grid aggregators (charging stations) with excessive demand due to the geographic location, peak time, etc. Existing solutions on demand response cannot achieve significant improvements based only on time-shifting the loads without considering the device properties such as charging modes and movement capabilities to enable geographic migration. Additionally, the information on the spare capacity at a few aggregators can aid in re-channeling the load from other aggregators facing excess demand to allow migration of devices. In this paper, we model these flexible properties of the devices as a mixed-integer non-linear problem (MINLP) to minimize excess load and the improve the utility (benefit) across all devices. We propose an online distributed low-complexity heuristic that prioritizes devices based on demand and deadlines to minimize the cumulative loss in utility. The proposed heuristic is tested on an exhaustive set of synthetic data and compared with solutions from a solver/optimization tool for the same runtime to show the impracticality of using a solver. A real-world EV testbed data is also tested with our proposed solution and other scheduling solutions to show the practicality of generating a feasible schedule and a loss improvement of at least 57.23%.

Analog computing using non-volatile memristors has emerged as a promising solution for energy-efficient deep learning. New materials, like perovskites-based memristors are recently attractive due to their cost-effectiveness, energy efficiency and flexibility. Yet, challenges in material diversity and immature fabrications require extensive experimentation for device development. Moreover, significant non-idealities in these memristors often impede them for computing. Here, we propose a synergistic methodology to concurrently optimize perovskite memristor fabrication and develop robust analog DNNs that effectively address the inherent non-idealities of these memristors. Employing Bayesian optimization (BO) with a focus on usability, we efficiently identify optimal materials and fabrication conditions for perovskite memristors. Meanwhile, we developed "BayesMulti", a DNN training strategy utilizing BO-guided noise injection to improve the resistance of analog DNNs to memristor imperfections. Our approach theoretically ensures that within a certain range of parameter perturbations due to memristor non-idealities, the prediction outcomes remain consistent. Our integrated approach enables use of analog computing in much deeper and wider networks, which significantly outperforms existing methods in diverse tasks like image classification, autonomous driving, species identification, and large vision-language models, achieving up to 100-fold improvements. We further validate our methodology on a 10$\times$10 optimized perovskite memristor crossbar, demonstrating high accuracy in a classification task and low energy consumption. This study offers a versatile solution for efficient optimization of various analog computing systems, encompassing both devices and algorithms.

Traditional robotic manipulation mostly focuses on collision-free tasks. In practice, however, many manipulation tasks (e.g., occluded object grasping) require the robot to intentionally collide with the environment to reach a desired task configuration. By enabling compliant robot motions, collisions between the robot and the environment are allowed and can thus be exploited, but more physical uncertainties are introduced. To address collision-rich problems such as occluded object grasping while handling the involved uncertainties, we propose a collision-inclusive planning framework that can transition the robot to a desired task configuration via roughly modeled collisions absorbed by Cartesian impedance control. By strategically exploiting the environmental constraints and exploring inside a manipulation funnel formed by task repetitions, our framework can effectively reduce physical and perception uncertainties. With real-world evaluations on both single-arm and dual-arm setups, we show that our framework is able to efficiently address various realistic occluded grasping problems where a feasible grasp does not initially exist.

Low-rank factorization is a popular model compression technique that minimizes the error $\delta$ between approximated and original weight matrices. Despite achieving performances close to the original models when $\delta$ is optimized, a performance discrepancy remains due to the separate optimization processes for low-rank factorization and model performance, resulting in unavoidable losses. We address this issue by introducing a novel joint optimization strategy for lossless low-rank weight factorization, which, for the first time, enhances the model's performance beyond the original. Our approach begins with a theoretical analysis of the relationship between low-rank factorization and model optimization objectives, establishing a precise perturbation range for matrix factorization errors on model performance. This challenge is then reformulated as a numerical rank deficiency problem with inequality constraints and develop a joint objective that simultaneously addresses factorization error and model performance. Based on the above analysis, we propose two optimization algorithms: \textbf{a lossless optimization algorithm} that maximizes model accuracy while ensuring compression, and \textbf{a compact optimization algorithm} that minimizes model size while preserving performance. These algorithms do not require fine-tuning and can directly compress numerous deep models to achieve lossless results. Our methods demonstrate robust efficacy across various vision and language tasks. For example, the compressed model reduced by 70\% on ResNext50 outperforms the original. Our code will be made public.

Evolving next-generation mobile networks is designed to provide ubiquitous coverage and networked sensing. With utility of multi-view sensing and multi-node joint transmission, cell-free is a promising technique to realize this prospect. This paper aims to tackle the problem of access point (AP) deployment in cell-free systems to balance the sensing accuracy and user rate. By merging the D-optimality with Euclidean criterion, a novel integrated metric is proposed to be the objective function for both max-sum and max-min problems, which respectively guarantee the overall and lowest performance in multi-user communication and target tracking scenario. To solve the corresponding high dimensional non-convex multi-objective problem, the Soft actor-critic (SAC) is utilized to avoid risk of local optimal result. Numerical results demonstrate that proposed SAC-based APs deployment method achieves $20\%$ of overall performance and $120\%$ of lowest performance.

We consider a bilevel learning framework for learning linear operators. In this framework, the learnable parameters are optimized via a loss function that also depends on the minimizer of a convex optimization problem (denoted lower-level problem). We utilize an iterative algorithm called `piggyback' to compute the gradient of the loss and minimizer of the lower-level problem. Given that the lower-level problem is solved numerically, the loss function and thus its gradient can only be computed inexactly. To estimate the accuracy of the computed hypergradient, we derive an a-posteriori error bound, which provides guides for setting the tolerance for the lower-level problem, as well as the piggyback algorithm. To efficiently solve the upper-level optimization, we also propose an adaptive method for choosing a suitable step-size. To illustrate the proposed method, we consider a few learned regularizer problems, such as training an input-convex neural network.

Training deep reinforcement learning (RL) agents necessitates overcoming the highly unstable nonconvex stochastic optimization inherent in the trial-and-error mechanism. To tackle this challenge, we propose a physics-inspired optimization algorithm called relativistic adaptive gradient descent (RAD), which enhances long-term training stability. By conceptualizing neural network (NN) training as the evolution of a conformal Hamiltonian system, we present a universal framework for transferring long-term stability from conformal symplectic integrators to iterative NN updating rules, where the choice of kinetic energy governs the dynamical properties of resulting optimization algorithms. By utilizing relativistic kinetic energy, RAD incorporates principles from special relativity and limits parameter updates below a finite speed, effectively mitigating abnormal gradient influences. Additionally, RAD models NN optimization as the evolution of a multi-particle system where each trainable parameter acts as an independent particle with an individual adaptive learning rate. We prove RAD's sublinear convergence under general nonconvex settings, where smaller gradient variance and larger batch sizes contribute to tighter convergence. Notably, RAD degrades to the well-known adaptive moment estimation (ADAM) algorithm when its speed coefficient is chosen as one and symplectic factor as a small positive value. Experimental results show RAD outperforming nine baseline optimizers with five RL algorithms across twelve environments, including standard benchmarks and challenging scenarios. Notably, RAD achieves up to a 155.1% performance improvement over ADAM in Atari games, showcasing its efficacy in stabilizing and accelerating RL training.