Sašo Karakatič

A survey of genetic algorithms for solving multi depot vehicle routing problem

We reviewed the use of genetic algorithms on the MDVRP (multi depot vehicle routing problem).Survey was made on every operator and setting of genetic algorithm for this problem.We tested different genetic operators and compared the results.We compared the genetic algorithms to other metaheuristic algorithms on MDVRP based on the results on standard benchmarks. This article presents a survey of genetic algorithms that are designed for solving multi depot vehicle routing problem. In this context, most of the articles focus on different genetic approaches, methods and operators, commonly used in practical applications to solve this well-known and researched problem. Besides providing an up-to-date overview of the research in the field, the results of a thorough experiment are presented and discussed, which evaluated the efficiency of different existing genetic methods on standard benchmark problems in detail. In this manner, the insights into strengths and weaknesses of specific methods, operators and settings are presented, which should help researchers and practitioners to optimize their solutions in further studies done with the similar type of the problem in mind. Finally, genetic algorithm based solutions are compared with other existing approaches, both exact and heuristic, for solving this same problem.

Improved classification with allocation method and multiple classifiers

We propose a new allocation method for building a classification ensemble.Allocation method uses multiple classifiers: the allocator and micro classifiers.Allocator separates the dataset and allocates them to one of micro classifiers.Allocator is based on one class SVM for anomaly detection.Results show improvement over basic classifiers and standard ensemble methods. Classification is the most used supervized machine learning method. As each of the many existing classification algorithms can perform poorly on some data, different attempts have arisen to improve the original algorithms by combining them. Some of the best know results are produced by ensemble methods, like bagging or boosting. We developed a new ensemble method called allocation. Allocation method uses the allocator, an algorithm that separates the data instances based on anomaly detection and allocates them to one of the micro classifiers, built with the existing classification algorithms on a subset of training data. The outputs of micro classifiers are then fused together into one final classification. Our goal was to improve the results of original classifiers with this new allocation method and to compare the classification results with existing ensemble methods. The allocation method was tested on 30 benchmark datasets and was used with six well known basic classification algorithms (J48, NaiveBayes, IBk, SMO, OneR and NBTree). The obtained results were compared to those of the basic classifiers as well as other ensemble methods (bagging, MultiBoost and AdaBoost). Results show that our allocation method is superior to basic classifiers and also to tested ensembles in classification accuracy and f-score. The conducted statistical analysis, when all of the used classification algorithms are considered, confirmed that our allocation method performs significantly better both in classification accuracy and f-score. Although the differences are not significant for each of the used basic classifier alone, the allocation method achieved the biggest improvements on all six basic classification algorithms. In this manner, allocation method proved to be a competitive ensemble method for classification that can be used with various classification algorithms and can possibly outperform other ensembles on different types of data.

Weighting and sampling data for individual classifiers and bagging with genetic algorithms

An imbalanced or inappropriate dataset can have a negative influence in classification model training. In this paper we present an evolutionary method that effectively weights or samples the tuples from the training dataset and tries to minimize the negative effects from innaprotirate datasets. The genetic algorithm with genotype of real numbers is used to evolve the weights or occurrence number for each learning tuple in the dataset. This technique is used with individual classifiers and in combination with the ensemble technique of bagging, where multiple classification models work together in a classification process. We present two variations - weighting the tuples and sampling the classification tuples. Both variations are experimentally tested in combination with individual classifiers (C4.5 and Naive Bayes methods) and in combination with bagging ensemble. Results show that both variations are promising techniques, as they produced better classification models than methods without weighting or sampling, which is also supported with statistical analysis.

Predicting Grades Based on Students’ Online Course Activities

We researched the possibility of predicting the final grades of university students with the help of online course management systems. By using the activity logs from the system we identify those variables that could be used during predictions. We experimentally narrowed-down the selection to two variables that would be useful for constructing linear regression models for grade prediction. The identified variables were the number of specific activities and the intermediate grades of the students. An experiment was conducted in order to evaluate the selection regarding five courses, which would show whether these two variables could help build a prediction model with accuracy of up to 91.7 % for a given course.

Evolving balanced decision trees with a multi-population genetic algorithm

Multi-population genetic algorithms have been used with success for several multi-objective optimization problems. In this paper, we present a new general multi-population genetic algorithm for evolving decision trees. It was designed to improve the possibility of evolving balanced decision trees, simultaneously optimized for the predictions of each class. Single-population genetic algorithms namely tend to construct decision trees with great variance in single class accuracies. The proposed approach is tested over 10 UCI datasets, and it is compared with a single-population genetic algorithm as well as with traditional decision-tree induction algorithms. Results show that the designed multi-population approach provides classification results comparable to C4.5 and CART in terms of accuracy and tree size, while outperforming them regarding balanced solutions (in terms of average class accuracy and range of single-class accuracies).

A Novel Approach to Generating Test Cases with Genetic Programming

Part of the automating software testing procedure includes the automation of test cases. Automation can lower the cost and effort and at the same time can increase the quality of test cases and consequently the testing procedure. Many different approaches for test case generation are available: generation from code, formal methods and different models, among others also from UML diagrams, more precisely from UML activity diagrams. Researchers use different techniques, of which genetic programming (GP) is very popular and was used in our research. In the proposed approach we generated test cases from the UML activity diagram, from which we constructed the binary decision tree structure, which is used as an instance in the evolution process of GP. The default tree structure is used throughout the whole evolution process, only the content (the testing parameters) of the nodes changes. The process of evolution consists of several genetic operators, such as selection, crossover and mutation. The main novelty of our method is a different fitness function than we can find in existing literature. In contrast to related work where the coverage is used - we used the error occurrence for our metric. The proposed method is demonstrated on the example of an automated teller machine (ATM), where we show how the full automation of test case generation and testing is a major advantage of our method.

Building boosted classification tree ensemble with genetic programming

Adaptive boosting (AdaBoost) is a method for building classification ensemble, which combines multiple classifiers built in an iterative process of reweighting instances. This method proves to be a very effective classification method, therefore it was the major part of our evolutionary inspired classification algorithm. In this paper, we introduce the Genetic Programming with AdaBoost (GPAB) which combines the induction of classification trees with genetic programming (GP) and AdaBoost for multiple class problems. Our method GPAB builds the ensemble of classification trees and uses AdaBoost through the evolution to weight instances and individual trees. To evaluate the potential of the proposed evolutionary method, we made an experiment where we compared the GPAB with Random Forest and AdaBoost on several standard UCI classification benchmarks. The results show that GPAB improves classification accuracy in comparison to other two classifiers.

Sensors and Functionalities of Non-Invasive Wrist-Wearable Devices: A Review

Wearable devices have recently received considerable interest due to their great promise for a plethora of applications. Increased research efforts are oriented towards a non-invasive monitoring of human health as well as activity parameters. A wide range of wearable sensors are being developed for real-time non-invasive monitoring. This paper provides a comprehensive review of sensors used in wrist-wearable devices, methods used for the visualization of parameters measured as well as methods used for intelligent analysis of data obtained from wrist-wearable devices. In line with this, the main features of commercial wrist-wearable devices are presented. As a result of this review, a taxonomy of sensors, functionalities, and methods used in non-invasive wrist-wearable devices was assembled.

Experiments with Lazy Evaluation of Classification Decision Trees Made with Genetic Programming.

In this paper, we present a lazy evaluation approach of classification decision trees with genetic programming. We describe and experiment with the lazy evaluation that does not evaluate the whole population but evaluates only the individuals that are chosen to participate in the tournament selection method. Further on, we used dynamic weights for the classification instances, that are linked to the chance of that instance getting picked for the evaluation process. These weights change based on the misclassification rate of the instance. We test our lazy evaluation approach on 10 standard classification benchmark datasets and show that not only lazy evaluation approach uses less time to evolve the good solution, but can even produce better solution due to changing instance weights and thus preventing the overfitting of the solutions.

Using similarity-based selection in evolutionary design of decision trees

When evaluating the process of building classification decision trees, it is necessary to assess the performance of constructed trees, as well as the speed and efficiency of the algorithm. Top-down induction algorithms are relatively simple and can quickly generate good solutions, however their deterministic nature often prevents them from finding globally optimal solutions. On the other hand, the evolutionary approach to decision tree building has yielded promising results by exploring and exploiting the entire search space. However, the standard evolutionary method of building decision trees uses the fitness-based selection of two trees for crossover, which can lead to premature convergence to a local, often sub-optimal solution. In order to maintain the diversity of the population over the course of evolution, we propose a novel method of selection that takes into consideration the similarity of trees in the crossover process, to prevent inbreeding. Several different approaches to evaluate the similarity between trees were designed and implemented. The approaches of both similar and diverse tree crossover were compared to the standard induction algorithm on twenty different data sets to determine the impact of similarity on the effectiveness and efficiency of the genetic algorithm.