In this paper, we study the quadratic assignment problem with a rank-one cost matrix (QAP-R1). Four integer-programming formulations are introduced of which three are assumed to have partial integer data. Unlike the standard quadratic assignment problem, some of our formulations can solve reasonably large instances of QAP-R1 with impressive running times and are faster than some metaheuristics. Pairwise relative strength of the LP relaxations of these formulations are also analyzed from theoretical and experimental points of view. Finally, we present a new metaheuristic algorithm to solve QAP-R1 along with its computational analysis. Our study offers the first systematic experimental analysis of integer-programming models and heuristics for QAP-R1. The benchmark instances with various characteristics generated for our study are made available to the public for future research work. Some new polynomially solvable special cases are also introduced. Summary of Contribution: This paper aims to advance our knowledge and ability in solving an important special case of the quadratic assignment problem. It shows how to exploit inherent properties of an optimization problem to achieve computational advantages, a strategy that was followed by researchers in model building and algorithm developments for decades. Our computational results attest to this time-tested general philosophy. The paper presents the first systematic computational study of the rank one quadratic assignment problem, along with new mathematical programming models and complexity analysis. We believe the theoretical and computational results of this paper will inspire further research on the topic and will be of significant value to practitioners using rank one quadratic assignment models.

The routing and wavelength assignment (RWA) problem is a classic and challenging problem in wavelength-division multiplexing (WDM) optical networks and has shown to be NP-hard. This paper studies the min-RWA problem with the objective of minimizing the number of required wavelengths and presents a new powerful neighborhood called Shift-and-Shaking (SAS). The proposed SAS integrates a high-level shift move to change the wavelength of one lightpath and two low-level ejection chain-based shaking (ECS) procedures to find the best routings for the related lightpaths. This new neighborhood is embedded into a simple iterated local search algorithm, called SAS-ILS, for solving min-RWA. The proposed SAS-ILS is tested on three sets of totally 113 widely studied instances in the literature. Comparison with other state-of-the-art algorithms shows that the SAS-ILS is able to improve 22 previous best known results, while matching the best known results for the remaining ones within short computational time.

The bipartite Boolean quadratic programming problem (BBQP) is a generalization of the well-studied NP-hard Boolean quadratic programming problem and can be regarded as a unified model for many graph theoretic optimization problems, including maximum weight-induced subgraph problems, maximum weight biclique problems, matrix factorization problems, and maximum cut problems on bipartite graphs. This paper introduces three main algorithms for solving the BBQP, based on three variants of tabu search, the first two consisting of strategic oscillation–tabu search (SO-TS) algorithms, which use destructive and constructive procedures to guide the search into unexplored and promising areas. The third algorithm, whichDoes also incorporates the SO-TS algorithms as solution improvement methods, uses a path relinking (PR) algorithm that is capable of further enhancing search performance. Experimental results demonstrate that all three algorithms perform very effectively compared with the best methods in the literature, and the PR algorithm joined with tabu search is able to discover new best solutions for two-thirds of the large problem instances and match the previous best known solutions for the other instances. Additional analysis discloses the contributions of the key ingredients of each of the proposed algorithms.

Semantic segmentation is a fundamental problem in computer vision. It is considered as a pixel-wise classification problem in practice, and most segmentation models use a pixel-wise loss as their optimization criterion. However, the pixel-wise loss ignores the dependencies between pixels in an image. Several ways to exploit the relationship between pixels have been investigated, \eg, conditional random fields (CRF) and pixel affinity based methods. Nevertheless, these methods usually require additional model branches, large extra memories, or more inference time. In this paper, we develop a region mutual information (RMI) loss to model the dependencies among pixels more simply and efficiently. In contrast to the pixel-wise loss which treats the pixels as independent samples, RMI uses one pixel and its neighbour pixels to represent this pixel. Then for each pixel in an image, we get a multi-dimensional point that encodes the relationship between pixels, and the image is cast into a multi-dimensional distribution of these high-dimensional points. The prediction and ground truth thus can achieve high order consistency through maximizing the mutual information (MI) between their multi-dimensional distributions. Moreover, as the actual value of the MI is hard to calculate, we derive a lower bound of the MI and maximize the lower bound to maximize the real value of the MI. RMI only requires a few extra computational resources in the training stage, and there is no overhead during testing. Experimental results demonstrate that RMI can achieve substantial and consistent improvements in performance on PASCAL VOC 2012 and CamVid datasets. The code is available at \url{https://github.com/ZJULearning/RMI}.