In current translation tasks, we discover that the weight distribution of attention has a fuzziness problem, specif-ically, its imperfection in detecting definitive information. In-tuitively, the most straightforward idea to alleviate fuzziness is to increase the number of attention heads at the expense of model complexities. However, our experiments prove this approach to be inefficient, and the effect is barely measurable. Instead, to tackle this problem, we explore an attention con-straint mechanism by introducing auxiliary attention. Instead of simply accumulating or averaging the base attention with auxiliary attention, we build a mathematical constraint function between them. The constraint function calibrates the original attention and further improves the model performance. Extensive experiments on neural machine translation tasks and abstract summarization tasks demonstrate that our attention constraint method can substantially outperform the strong baselines. The exact mathematical formulas and experimental analysis reveal the influence of constraint function on attention: attention tends to assign more weights to the decisive inputs, becoming sparse and sharp. Our code is available on Github11https://github.com/auxiliaryattention/AttentionConstraintMechanism.

Previous approaches on association rule mining in recommendation have already achieved promising performances. However, to the best of our knowledge, they seldom simultaneously take the profit and placement factor into consideration. In E-commerce recommendation scenario, the order of the recommendation reflects as placement. In this paper, we propose a novel placement-and-profit-aware association rule mining algorithm to maximize profit as well as maintaining recommendation accuracy. We also propose two metrics: Expectation of Profit (EOP), which measures the overall profit, and Expectation of Click rate (EOC), which measures the user experience. Experiments on SPMF dataset show that the proposed algorithm can improve the EOP significantly with only slight decrease in EOC.

Cognitive Radio Network (CRN) technology allows secondary users (SUs) to transmit data exploiting the wireless resources not utilized by licensed primary users (PUs). Channel exploration by SUs for finding transmission opportunities incurs non-negligible costs and is a key challenge in successful operation of CRN. In this paper, we investigate the joint effect of channel exploration order and stopping rule for channel selection with the goal of maximizing SUs' throughput. The main contribution of this paper is the near optimal exploration order and optimal stopping rule, paired with a new and more general characterization of PUs' traffic configurations in CRN. The proposed algorithm uses an analytical model that incorporates channel and PUs' activity history to make the best decision at each decision point, according to the current SUs' exploration results. The simulation study shows that the proposed algorithm significantly improves the SUs' throughput over the existing methods. The proposed optimal stopping algorithm runs in polynomial-time, which is another major contribution of this paper, since it is a significant complexity reduction from the typical exponential order of backward induction algorithms.

In Cognitive Radio Networks (CRNs) secondary users (SUs) are allowed to transmit data exploiting the wireless resources (i.e., channels) not utilized by licensed primary users (PUs). Efficient resource discovery for finding available channels is a key aspect for successful SUs' operation. In this paper we investigate and propose different algorithms to discover the best CRN channels that, exploited for transmission jointly as aggregate OFDM carriers, will maximize SUs' throughput. By using CRN channels statistics such as channel availability and link quality, together with current channel discovery results, SUs attempt to make the best choice at each decision point. This is choosing to continue the discovery process or stop and transmit using the aggregated resources found available so far. All the proposed algorithms run in polynomial time, which is a significant complexity reduction from the typical exponential order of optimal stopping rule algorithms. The simulation study confirms the validity of our algorithms in different CRN scenarios, comparing their throughput performance with the theoretically best achievable results.

Cognitive Radio Networks (CRNs) aim to maximize the utilization of existing wireless channels by allowing secondary users (SUs) to transmit when licensed primary users (PUs) are not using the same channels. An SU monitors the CRN channels, sensing PU presence to avoid interference and estimating the link quality before transmitting. It stops when one or more available channels with satisfactory link quality are found. Algorithms for making the optimal decision regarding when to stop exploring the channels and start transmitting are expensive in terms of time and space, which are both scarce in hardware-constrained SUs, such as mobile devices. In this paper, we propose a low-complexity algorithm, which utilizes link quality and PU-activity statistics of the CRN channels to pre-compute a set of decision thresholds that will aid the channel exploration phase in maximizing SU-throughput. Our algorithm takes quadratic time and space for offline computations and linear time and space for online processing, which makes it very suitable for space and energy constrained mobile SUs. Our extensive simulation study and analytical model matching the simulation results demonstrate our solution's validity by showing the closeness of throughput and delay performances with the optimum solution as well as solutions by the well-known backward induction method, which often runs in exponential time for offline computations.

This paper considers the channel-assignment and scheduling in wireless mesh networks that employ multiple radios and multiple channels. In contrast to the various algorithms available in the literature, we explicitly model the delay overhead that is incurred during channel switching, and use that delay in the design of algorithms. We prove that the well known Greedy Maximal Scheduling (GMS) algorithm does not have any provable efficiency ratio when the switching overhead is considered. We present a centralized algorithm (CGSSO), and a dynamic algorithm (DMSSO), both of which consider switching overhead. Simulation results show that the proposed algorithms significantly outperform other algorithms in packet throughput and average packet delay metrics. Results also show that the improvements in performance become more pronounced as the switching delay increases.

In this paper, we consider the important aspect of quality of service (QoS) in wireless mesh networks, focusing on packet delays and packet drops. We observe that the options of solely focusing on throughput, and of only depending on the QoS type characterization by the application level protocol is not sufficient. We propose the use of multiple queues to hold the packets based on their QoS requirement (not necessarily corresponding to the QoS application types). We suggest the technique of reinforcement learning (RL) for the important step of assigning packet to one of the queues, and present an implementation using the TD(0) algorithm. We present various simulation results comparing our algorithm to other known algorithms and frameworks. Our results indicate that our RL-based algorithm is significantly better than previously known results in packet drop ratio.

This paper considers the problem of maximizing the system throughput in IEEE 802.16 broadband access networks with mesh topology, and the following results are presented. We consider a linear chain network and discuss its applicability for providing cost effective solutions in sparsely populated areas, such as interstate highways and rural communities. We provide an optimal scheduling algorithm and establish an analytical result on the length of the schedule for linear chain networks. We also consider the problem of routing and packet scheduling in general topology, and show its NP-completeness. Based on our optimal algorithm for linear networks, we propose algorithms that find routes and schedules of packet transmissions in general mesh topologies. The performance of our proposed algorithms is analyzed using the NS-2 simulator. The results show that the proposed algorithms perform significantly better than other existing algorithms.

This paper studies the effects of the sub-carrier grouping on the scheduling problem in a wireless network with multiple carriers. In particular, the authors consider the trade-offs between the size of the sub-carrier groups and the link throughput for various sub-carrier grouping models, with each corresponding to a different level of CSI tracking and system design complexity. The paper assumes that a single user is communicating with the base station using the entire sub-carriers. This assumption is motivated by the OFDM, which uses multiple carriers in order to benefit from the frequency selective fading introduced by spreading the information bits over the entire signal band. An offline dynamic programming algorithm for finding an optimal solution for the case of spectrum- and time-perfect estimates is also presented. The performance can potentially be improved by using high-level modulation on the sub-carriers with good channel conditions.

The design of logical topologies in wavelength-routing multihop optical networks is a well-studied problem. We consider logical topology (LT) design over the popular ring and linear topologies. Our objective is the minimization of the electronic processing delay for the worst case traffic flow. For uniform traffic between nodes, this delay minimization corresponds to minimizing the number of hops on a shortest path between the farthest two nodes in the logical topology (the diameter of the logical topology). The simple structure of the physical topologies enables us to present a rigorous analysis of the problem. We present lower bounds for the achievable diameter wherever possible and propose practical logical topology design algorithms and corresponding upper bounds. We also present an application of the LT designs in the linear topology to the survivability of ring networks.

The blocking performance of wavelength routing WDM optical networks can be improved by employing wavelength conversion. In this paper, we address the problem of optimally placing a limited number of wavelength converters in mesh topologies. Two objective functions, namely, minimizing the average blocking probability and minimizing the maximum blocking probability over all routes, are considered. We first extend an earlier analytical model to compute the blocking probability on an arbitrary route in a mesh topology, given the traffic and locations of converters. We then propose heuristic algorithms to place wavelength converters, and evaluate the performance of the proposed heuristics using the analytical model.