On the differential evolution algorithm and its appliation to training radial basis function networks

The parameter setting of a differential evolution algorithm must meet several requirements: efficiency, effectiveness, and reliability. Problems vary. The solution of a particular problem can be represented in different ways. An algorithm most efficient in dealing with a particular representation may be less efficient in dealing with other representations. The development of differential evolution-based methods contributes substantially to research on evolutionary computing and global optimization in general. The objective of this study is to investigatethe differential evolution algorithm, the intelligent adjustment of its controlparameters, and its application. In the thesis, the differential evolution algorithm is first examined using different parameter settings and test functions. Fuzzy control is then employed to make control parameters adaptive based on an optimization process and expert knowledge. The developed algorithms are applied to training radial basis function networks for function approximation with possible variables including centers, widths, and weights of basis functions and both having control parameters kept fixed and adjusted by fuzzy controller. After the influence of control variables on the performance of the differential evolution algorithm was explored, an adaptive version of the differential evolution algorithm was developed and the differential evolution-based radial basis function network training approaches were proposed. Experimental results showed that the performance of the differential evolution algorithm is sensitive to parameter setting, and the best setting was found to be problem dependent. The fuzzy adaptive differential evolution algorithm releases the user load of parameter setting and performs better than those using all fixedparameters. Differential evolution-based approaches are effective for training Gaussian radial basis function networks.

Modelling non-industrial private forest landowners' strategic decision making by using logistic regression and neural networks: case of predicting the choice of forest taxation basis

Summary

Use of neural networks in robot positioning of large flexible redundant manipulators

Deflection compensation of flexible boom structures in robot positioning is usually done using tables containing the magnitude of the deflection with inverse kinematics solutions of a rigid structure. The number of table values increases greatly if the working area of the boom is large and the required positioning accuracy is high. The inverse kinematics problems are very nonlinear, and if the structure is redundant, in some cases it cannot be solved in a closed form. If the structural flexibility of the manipulator arms is taken into account, the problem is almost impossible to solve using analytical methods. Neural networks offer a possibility to approximate any linear or nonlinear function. This study presents four different methods of using neural networks in the static deflection compensation and inverse kinematics solution of a flexible hydraulically driven manipulator. The training information required for training neural networks is obtained by employing a simulation model that includes elasticity characteristics. The functionality of the presented methods is tested based on the simulated and measured results of positioning accuracy. The simulated positioning accuracy is tested in 25 separate coordinate points. For each point, the positioning is tested with five different mass loads. The mean positioning error of a manipulator decreased from 31.9 mm to 4.1 mm in the test points. This accuracy enables the use of flexible manipulators in the positioning of larger objects. The measured positioning accuracy is tested in 9 separate points using three different mass loads. The mean positioning error decreased from 10.6 mm to 4.7 mm and the maximum error from 27.5 mm to 11.0 mm.

Forecasting of wind speeds and directions with artificial neural networks

In this master’s thesis, wind speeds and directions were modeled with the aim of developing suitable models for hourly, daily, weekly and monthly forecasting. Artificial Neural Networks implemented in MATLAB software were used to perform the forecasts. Three main types of artificial neural network were built, namely: Feed forward neural networks, Jordan Elman neural networks and Cascade forward neural networks. Four sub models of each of these neural networks were also built, corresponding to the four forecast horizons, for both wind speeds and directions. A single neural network topology was used for each of the forecast horizons, regardless of the model type. All the models were then trained with real data of wind speeds and directions collected over a period of two years in the municipal region of Puumala in Finland. Only 70% of the data was used for training, validation and testing of the models, while the second last 15% of the data was presented to the trained models for verification. The model outputs were then compared to the last 15% of the original data, by measuring the mean square errors and sum square errors between them. Based on the results, the feed forward networks returned the lowest generalization errors for hourly, weekly and monthly forecasts of wind speeds; Jordan Elman networks returned the lowest errors when used for forecasting of daily wind speeds. Cascade forward networks gave the lowest errors when used for forecasting daily, weekly and monthly wind directions; Jordan Elman networks returned the lowest errors when used for hourly forecasting. The errors were relatively low during training of the models, but shot up upon simulation with new inputs. In addition, a combination of hyperbolic tangent transfer functions for both hidden and output layers returned better results compared to other combinations of transfer functions. In general, wind speeds were more predictable as compared to wind directions, opening up opportunities for further research into building better models for wind direction forecasting.

Pretraining Convolutional Neural Networks for Visual Recognition

Convolutional Neural Networks (CNN) have become the state-of-the-art methods on many large scale visual recognition tasks. For a lot of practical applications, CNN architectures have a restrictive requirement: A huge amount of labeled data are needed for training. The idea of generative pretraining is to obtain initial weights of the network by training the network in a completely unsupervised way and then fine-tune the weights for the task at hand using supervised learning. In this thesis, a general introduction to Deep Neural Networks and algorithms are given and these methods are applied to classification tasks of handwritten digits and natural images for developing unsupervised feature learning. The goal of this thesis is to find out if the effect of pretraining is damped by recent practical advances in optimization and regularization of CNN. The experimental results show that pretraining is still a substantial regularizer, however, not a necessary step in training Convolutional Neural Networks with rectified activations. On handwritten digits, the proposed pretraining model achieved a classification accuracy comparable to the state-of-the-art methods.

Deep convolutional neural networks for semantic video object segmentation

In this thesis, we propose to infer pixel-level labelling in video by utilising only object category information, exploiting the intrinsic structure of video data. Our motivation is the observation that image-level labels are much more easily to be acquired than pixel-level labels, and it is natural to find a link between the image level recognition and pixel level classification in video data, which would transfer learned recognition models from one domain to the other one. To this end, this thesis proposes two domain adaptation approaches to adapt the deep convolutional neural network (CNN) image recognition model trained from labelled image data to the target domain exploiting both semantic evidence learned from CNN, and the intrinsic structures of unlabelled video data. Our proposed approaches explicitly model and compensate for the domain adaptation from the source domain to the target domain which in turn underpins a robust semantic object segmentation method for natural videos. We demonstrate the superior performance of our methods by presenting extensive evaluations on challenging datasets comparing with the state-of-the-art methods.

Spatial modeling techniques for mapping and characterization of acid sulfate soils

Acid sulfate (a.s.) soils constitute a major environmental issue. Severe ecological damage results from the considerable amounts of acidity and metals leached by these soils in the recipient watercourses. As even small hot spots may affect large areas of coastal waters, mapping represents a fundamental step in the management and mitigation of a.s. soil environmental risks (i.e. to target strategic areas). Traditional mapping in the field is time-consuming and therefore expensive. Additional more cost-effective techniques have, thus, to be developed in order to narrow down and define in detail the areas of interest. The primary aim of this thesis was to assess different spatial modeling techniques for a.s. soil mapping, and the characterization of soil properties relevant for a.s. soil environmental risk management, using all available data: soil and water samples, as well as datalayers (e.g. geological and geophysical). Different spatial modeling techniques were applied at catchment or regional scale. Two artificial neural networks were assessed on the Sirppujoki River catchment (c. 440 km2) located in southwestern Finland, while fuzzy logic was assessed on several areas along the Finnish coast. Quaternary geology, aerogeophysics and slope data (derived from a digital elevation model) were utilized as evidential datalayers. The methods also required the use of point datasets (i.e. soil profiles corresponding to known a.s. or non-a.s. soil occurrences) for training and/or validation within the modeling processes. Applying these methods, various maps were generated: probability maps for a.s. soil occurrence, as well as predictive maps for different soil properties (sulfur content, organic matter content and critical sulfide depth). The two assessed artificial neural networks (ANNs) demonstrated good classification abilities for a.s. soil probability mapping at catchment scale. Slightly better results were achieved using a Radial Basis Function (RBF) -based ANN than a Radial Basis Functional Link Net (RBFLN) method, narrowing down more accurately the most probable areas for a.s. soil occurrence and defining more properly the least probable areas. The RBF-based ANN also demonstrated promising results for the characterization of different soil properties in the most probable a.s. soil areas at catchment scale. Since a.s. soil areas constitute highly productive lands for agricultural purpose, the combination of a probability map with more specific soil property predictive maps offers a valuable toolset to more precisely target strategic areas for subsequent environmental risk management. Notably, the use of laser scanning (i.e. Light Detection And Ranging, LiDAR) data enabled a more precise definition of a.s. soil probability areas, as well as the soil property modeling classes for sulfur content and the critical sulfide depth. Given suitable training/validation points, ANNs can be trained to yield a more precise modeling of the occurrence of a.s. soils and their properties. By contrast, fuzzy logic represents a simple, fast and objective alternative to carry out preliminary surveys, at catchment or regional scale, in areas offering a limited amount of data. This method enables delimiting and prioritizing the most probable areas for a.s soil occurrence, which can be particularly useful in the field. Being easily transferable from area to area, fuzzy logic modeling can be carried out at regional scale. Mapping at this scale would be extremely time-consuming through manual assessment. The use of spatial modeling techniques enables the creation of valid and comparable maps, which represents an important development within the a.s. soil mapping process. The a.s. soil mapping was also assessed using water chemistry data for 24 different catchments along the Finnish coast (in all, covering c. 21,300 km2) which were mapped with different methods (i.e. conventional mapping, fuzzy logic and an artificial neural network). Two a.s. soil related indicators measured in the river water (sulfate content and sulfate/chloride ratio) were compared to the extent of the most probable areas for a.s. soils in the surveyed catchments. High sulfate contents and sulfate/chloride ratios measured in most of the rivers demonstrated the presence of a.s. soils in the corresponding catchments. The calculated extent of the most probable a.s. soil areas is supported by independent data on water chemistry, suggesting that the a.s. soil probability maps created with different methods are reliable and comparable.

Application of Neural Network Methods in the Prediction of Stock Markets

The purpose of the research is to define practical profit which can be achieved using neural network methods as a prediction instrument. The thesis investigates the ability of neural networks to forecast future events. This capability is checked on the example of price prediction during intraday trading on stock market. The executed experiments show predictions of average 1, 2, 5 and 10 minutes’ prices based on data of one day and made by two different types of forecasting systems. These systems are based on the recurrent neural networks and back propagation neural nets. The precision of the predictions is controlled by the absolute error and the error of market direction. The economical effectiveness is estimated by a special trading system. In conclusion, the best structures of neural nets are tested with data of 31 days’ interval. The best results of the average percent of profit from one transaction (buying + selling) are 0.06668654, 0.188299453, 0.349854787 and 0.453178626, they were achieved for prediction periods 1, 2, 5 and 10 minutes. The investigation can be interesting for the investors who have access to a fast information channel with a possibility of every-minute data refreshment.