While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by 17.9% and 7.3% on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY.

Erbium-doped fiber amplifier (EDFA) is an optical amplifier/repeater device used to boost the intensity of optical signals being carried through fiber optic communication networks. A highly accurate EDFA model - to predict the signal gain for each channel - is required because of its crucial role in optical network management and optimization. EDFA channel inputs (i.e. features) either carry signal or are idle, therefore they can be treated as binary features. However, channel outputs (and the corresponding signal gains) are continuous values. Labeled training data is very expensive to collect for EDFA devices, therefore we devise an active learning strategy suitable for binary features to overcome this issue. We propose to take advantage of sparse linear models to simplify the predictive model. This approach improves signal gain prediction and accelerates active learning query generation. We show the performance of our proposed active learning strategies on simulated data and real EDFA data.

New emerging high-dimensional data sets have made traditional clustering algorithms increasingly inefficient. More sophisticated approaches are required to cope with the increasing dimensionality and cardinality of such data sets. Feature selection methods are proposed as a solution to deal with this problem, however they fail for data sets where the attribute support for different clusters is not the same. For this category of data sets subspace clustering algorithms have been introduced over the past decade. We approach this problem from the perspective of Genetic Algorithms by adopting a hierarchical data structure deployed in three stages. 1) a traditional clustering algorithm is applied independently to each attribute of the data set, thus defining a grid of potential 1-d cluster centroids. 2) representing multi-dimensional cluster centroids by indexing 1-d cluster centroids. 3) converting the problem of finding the best combination of cluster centroids into that of discrete optimization and applying a multi-objective evolutionary algorithm, which uses group fitness evaluation to give a fitness to a group of clusters, as defined by process 2. Synthetic data sets with different characteristics are generated as the ground truth to evaluate the resulting algorithm for Evolutionary Subspace Clustering (ESC) as well as benchmark against alternative subspace and full-space clustering algorithms. ESC returns competitive accuracy and while typically utilizing less attributes and scaling as attribute count increases.

The ultimate goal of subspace clustering algorithms is to identify both the subset of attributes supporting a cluster and the location of the cluster in the subspace. In this work a generic evolutionary approach to bottom-up subspace clustering is proposed consisting of three steps. The first applies a non-evolutionary clustering algorithm attribute-wise to establish the lattice from which subspace clusters will be designed. In the second step a multi-objective Genetic Algorithm (MOGA) is used to evolve good candidate subspace clusters (CSC) through a combinatorial search w.r.t. the attribute-wise lattice from step 1. The third step then searches in the space of CSC from the population of the the first MOGA to find the best combination of subspace clusters, again under a MOGA formulation. Important properties of the approach are that a standard clustering algorithm is deployed in step one to build the initial lattice of attribute-wise clusters. This helps to decouple the computational expense of clustering using Evolutionary Computation, with the MOGA applied in steps 2 and 3 building clusters through a combinatorial search relative to the original lattice parameters. Benchmarking on data sets with tens to hundreds of attributes illustrates the feasibility of the approach.

Subspace clustering algorithms in their most general form attempt to describe data with clusters that are not constrained to index a common set of attributes. Previous evolutionary approaches to this problem have assumed a weaker model in which clusters are built in a common subset. Moreover, a filter method is generally assumed in which a classical clustering algorithm is employed in the inner loop. Needless to say, this presents a considerable computational overhead. In this work we recognize the utility of assuming a `bottom-up' approach to subspace clustering. Specifically, we apply a classical clustering algorithm to each attribute to establish 1-d clusters that are then indexed by a MOGA to design a population of subspace clusters. The ensuing search is entirely in terms of a combinatorial optimization problem, thus computationally very efficient. A final single objective GA is then applied to search the set of subspace clusters identified under the MOGA for the most suitable combination.

For a mobile robot to move in a known environment and operate successfully, first it needs to robustly determine its initial position and orientation relative to the map, and then update its position while moving in the environment. Thus determining robot’s position is one of the most important tasks in mobile robotics. This task consists of “global localization‟ and “robot’s pose tracking‟. In this paper two recent sample-based evolutionary methods for globally localizing the position of a mobile robot are proposed. The first method is a modified version of genetic algorithm called Differential Evolution (DE) which is based on natural selection. The second one is Particle Swarm Optimization (PSO) which is based on bird flocking. DE evaluates initial population using the probabilistic motion and observation models and the evolution of the individuals is performed by evolutionary operators. PSO adjusts the velocity and location of particles towards target (robot’s pose) through a problem space on the basis of information about each particle’s previous best location and the best previous location of its neighbors. Our results illustrate the excellence of these two methods over standard Monte Carlo localization algorithm with regard to convergence rate, speed and computational cost.