Mehmet Fatih Amasyali

Classifier Ensembles with the Extended Space Forest

The extended space forest is a new method for decision tree construction in which training is done with input vectors including all the original features and their random combinations. The combinations are generated with a difference operator applied to random pairs of original features. The experimental results show that extended space versions of ensemble algorithms have better performance than the original ensemble algorithms. To investigate the success dynamics of the extended space forest, the individual accuracy and diversity creation powers of ensemble algorithms are compared. The Extended Space Forest creates more diversity when it uses all the input features than Bagging and Rotation Forest. It also results in more individual accuracy when it uses random selection of the features than Random Subspace and Random Forest methods. It needs more training time because of using more features than the original algorithms. But its testing time is lower than the others because it generates less complex base learners.

Predictions of oil/chemical tanker main design parameters using computational intelligence techniques

Abstract: Ship design and construction are complicated and expensive processes. In the pre-design stage, before the construction according to some special rules, determination of the main ship parameters is very important. In this study, instead of traditional methods, the oil/chemical tanker main design parameters are estimated by 18 computational intelligence methods. Therefore, all the data of 114 tankers in operation are used in the experiments in order to estimate a parameter from the remaining ones. Main result of this article is to show that, except for the speed parameter, the main parameters of tankers can be estimated sufficiently well for pre-design stage without having to apply conventional but arduous ship modeling experiments.

Locally adaptive k parameter selection for nearest neighbor classifier: one nearest cluster

The k nearest neighbors (k-NN) classification technique has a worldly wide fame due to its simplicity, effectiveness, and robustness. As a lazy learner, k-NN is a versatile algorithm and is used in many fields. In this classifier, the k parameter is generally chosen by the user, and the optimal k value is found by experiments. The chosen constant k value is used during the whole classification phase. The same k value used for each test sample can decrease the overall prediction performance. The optimal k value for each test sample should vary from others in order to have more accurate predictions. In this study, a dynamic k value selection method for each instance is proposed. This improved classification method employs a simple clustering procedure. In the experiments, more accurate results are found. The reasons of success have also been understood and presented.

Clustering Application Benchmark

An application benchmark based on a set of clustering algorithms is described in this paper. The details of algorithms (K-means online, K-means batch, SOM-1 dimension, SOM-2 dimension, hierarchical K-means online and hierarchical SOM-1 dimension) are given. The code provided complies with ANSI C specifications, as a result is highly portable. The benchmark has been tested on various platforms using different compilers

KNN parameter selection via meta learning

In this study, the K Nearest Neighbors parameter k is predicted by system. Meta learning method is used for prediction. Getting training set with meta-features, 200 data sets were used. For each of them, 16 meta-features were extracted. The K Nearest Neighbour algorithm was applied each of them with most common 6 k values the best one is selected. With this training set it is possible to predict a new data sets best k value. In 200 data sets the most common k value which has best performance is 1. 4 methods are applied on the model. Generally all methods used same features and some meta-features are never used.

Improved Space Forest: A Meta Ensemble Method

The performance of the ensemble algorithms is related with the individual accuracy of the base learners and their results diversity. Individual accuracy of a base learner is directly related to the similarity between the original training set and the base learner’s training set. When a modified training set by randomly selecting features/classes/samples is given to the base learners, the diversity is created but the individual accuracy is decreased. From this point of view, different ensemble algorithms can be seen as a selection between having more accurate but less diverse base learners and having more diverse but less accurate base learners. We propose a meta ensemble method named as improved space forest which adds generated and (hopefully) more accurate features to the original features. The new features are obtained from randomly selected original features. When the new features are more distinctive than the original ones, they are selected by the learners. So, the ensemble may have more accurate base learners. However, a different improved space is generated for each learner to create diversity. The proposed method can be used with different ensemble methods. We compared original and improved space versions of bagging, random forest, and rotation forest algorithms. Improved space versions have generally better or comparable results than the original ones. We also present a theoretical framework to analyze the individual accuracies and diversities of the base learners.

Ensemble learning with surrogate splits

Surrogate splits are used to classify test samples having missing values. In this work, they are used to produce different decisions from the same decision tree. In the popular ensemble algorithms, different sub-samples and sub-spaces are used to produce different decisions. But, in our approach, different versions of a test sample are generated by randomly deleting some features. For each version of the test sample, a different decision can be generated by using surrogate splits. 41 UCI datasets are used to compare original and surrogate split versions of the ensemble algorithms. Surrogate split versions have generally better performance than the original ones. The proposed method can be used within any ensemble algorithm using decision trees as its base learner.

A fast 3D exploration algorithm for autonomous aerial robots

In the scope of the study, it was aimed to discover a closed multi-storey environment with autonomous air robots and produce a three dimensional map. In order to reduce the complexity of the three-dimensional exploration algorithm, we have developed Target Elimination Method that can calculate the result by narrowing the problem space. The Target Elimination Method ensures that the exploration algorithm is faster to explore the environment as it reduces the calculation time.

An architecture for multi-robot localization and mapping in the Gazebo/Robot Operating System simulation environment:

Robots are an important part of urban search and rescue tasks. World wide attention has been given to developing capable physical platforms that would be beneficial for rescue teams. It is evident that use of multi-robots increases the effectiveness of these systems. The Robot Operating System (ROS) is becoming a standard platform for the robotics research community for both physical robots and simulation environments. Gazebo, with connectivity to the ROS, is a three-dimensional simulation environment that is also becoming a standard. Several simultaneous localization and mapping algorithms are implemented in the ROS; however, there is no multi-robot mapping implementation. In this work, two multi-robot mapping algorithm implementations are presented, namely multi-robot gMapping and multi-robot Hector Mapping. The multi-robot implementations are tested in the Gazebo simulation environment. Also, in order to achieve a more realistic simulation, every incremental robot movement is modeled with rotational and translational noise.

Text2arff: A text representation library

Which features are the most important for the text classification tasks? In the automatic text categorization area, several studies seek answers to this question. In this paper, new version of Text2arff (a library for text representation) and its new features (word2vec, Word trajectories, etc.) are presented. Also, the software is now a java library which can be used in the users own projects. In the experiments, the library is run on two sample datasets. The results show that the effect of text representation method is bigger than the classification method. This result also emphasizes the importance of developing new test representation methods.