This paper proposes a variant of multiple-play stochastic bandits tailored to resource allocation problems arising from LLM applications, edge intelligence, etc. The model is composed of finite number of arms and plays. Each arm has a stochastic number of capacities, and each unit of capacity is associated with a reward function. Each play is associated with a priority weight. When multiple plays compete for the arm capacity, the arm capacity is allocated in a larger priority weight first manner. Instance independent and instance dependent regret lower bounds are proved, revealing the impact of model parameters on the hardness of learning the optimal allocation policy. When model parameters are given, we design an algorithm named MSB-PRS-OffOpt to locate the optimal play allocation policy with a polynomial computational complexity in the number of arms and plays. Utilizing MSB-PRS-OffOpt as a subroutine, an approximate upper confidence bound (UCB) based algorithm is designed, which has instance independent and instance dependent regret upper bounds matching the corresponding lower bound up to acceptable factors. To this end, we address nontrivial technical challenges arising from optimizing and learning under a special nonlinear combinatorial utility function induced by the prioritized resource sharing mechanism.

The end-to-end image dehazing network relies on paired training data. However, there is limited training data available for real non-homogeneous dehazing, which limits the performance of dehazing networks on real non-homogeneous hazy images. To overcome this limitation, we propose a method to augment the training datasets for real non-homogeneous image dehazing. Different with existing methods that introduce augment data from other datasets, we estimate the degradations and synthesize additional hazy images by applying it to other scenes in the same dataset. During the estimation, we proposed a new degradation model based on 3D attenuation coefficient for describing non-homogeneous degradation. To solve the 3D attenuation coefficient, we propose a clue kernel to overcome the scene-dependence of the estimated degradation. The experimental results show that our proposed method can effectively augment the training dataset and improve the performance and robustness of the dehazing network. Compared to state-of-the-art methods, our approach outperforms the best performing method by 4.65% SSIM on the Dense-NH-HAZE dataset.

It is a longstanding problem that Q-learning suffers from the overestimation bias. This issue originates from the fact that Q-learning uses the expectation of maximum Q-value to approximate the maximum expected Q-value. A number of algorithms, such as Double Q-learning, were proposed to address this problem by reducing the estimation of maximum Q-value, but this may lead to an underestimation bias. Note that this underestimation bias may have a larger performance penalty than the overestimation bias. Different from previous algorithms, this article studies this issue from a fresh perspective, i.e., meta-learning view, which leads to our Meta-Debias Q-learning. The main idea is to extract the maximum expected Q-value with meta-learning over multiple tasks to remove the estimation bias of maximum Q-value and help the agent choose the optimal action more accurately. However, there are two challenges: (1) How to automatically select suitable training tasks? (2) How to positively transfer the meta-knowledge from selected tasks to remove the estimation bias of maximum Q-value? To address the two challenges mentioned above, we quantify the similarity between the training tasks and the test task. This similarity enables us to select appropriate “partial” training tasks and helps the agent extract the maximum expected Q-value to remove the estimation bias. Extensive experiment results show that our Meta-Debias Q-learning outperforms SOTA baselines drastically in three evaluation indicators, i.e., maximum Q-value, policy, and reward. More specifically, our Meta-Debias Q-learning only underestimates \(1.2*10^{-3}\) than the maximum expected Q-value in the multi-armed bandit environment and only differs \(5.04\%-5\%=0.04\%\) than the optimal policy in the two states MDP environment. In addition, we compare the uniform weight and our similarity weight. Experiment results reveal fundamental insights into why our proposed algorithm outperforms in the maximum Q-value, policy, and reward.