Sean Van Osselaer
Murdoch University, Western Australia
This project paper refers to experiments towards the classification of Iris plants with back propagation neural networks (BPNN). The problem concerns the identification of Iris plant species on the basis of plant attribute measurements. The paper outlines background information concerning the problem, making reference to statistics and value constraints identified in the course of the project. There is an outline of the algorithm of techniques used within the project, with descriptions of these techniques and their context. A discussion concerning the experimental setup is included, describing the implementation specifics of the project, preparatory actions, and the experimental results. The results generated by the networks constructed are presented, with the results being discussed and compared towards identification of the fittest architecture for the problem constrained by the data set. In conclusion, the fittest architecture is identified, and a justification concerning its selection offered.
Keywords : Iris, back propagation neural network, BPNN
This project paper is related to the use of back propagation neural networks (BPNN) towards the identification of iris plants on the basis of the following measurements: sepal length, sepal width, petal length, and petal width. There is a comparison of the fitness of neural networks with input data normalised by column, row, sigmoid, and column constrained sigmoid normalisation. Also contained within the paper is an analysis of the performance results of back propagation neural networks with various numbers of hidden layer neurons, and differing number of cycles (epochs). The analysis of the performance of the neural networks is based on several criteria: incorrectly identified plants by training set (recall) and testing set (accuracy), specific error within incorrectly identified plants, overall data set error as tested, and class identification precision. The fittest network architecture identified used column normalisation, 40000 cycles, 1 hidden layer with 9 hidden layer neurons, a step width of 0.15, a maximum non-propagated error of 0.1, and a value of 1 for the number of update steps.
This project makes use of the well known Iris dataset, which refers to 3 classes of 50 instances each, where each class refers to a type of Iris plant. The first of the classes is linearly distinguishable from the remaining two, with the second two not being linearly separable from each other. The 150 instances, which are equally separated between the 3 classes, contain the following four numeric attributes: sepal length and width, petal length and width. A sepal is a division in the calyx, which is the protective layer of the flower in bud, and a petal is the divisions of the flower in bloom. The minimum values for the raw data contained in the data set are as follows (measurements in centimetres): sepal length (4.3), sepal width (2.0), petal length (1.0), and petal width (0.1). The maximum values for the raw data contained in the data set are as follows (measurements in centimetres): sepal length (7.9), sepal width (4.4), petal length (6.9), and petal width (2.5). In addition to these numeric attributes, each instance also includes an identifying class name, each of which is one of the following: Iris Setosa, Iris Versicolour, or Iris Virginica.
ALGORITHM OF TECHNIQUE USE
Data set construction
This project uses a two data set approach. The first of these sets is the training set, which is used for the actual training of the network, and for the determination of the networks recall ability. The second data set is the testing data set, which is not used in the training process, and is used to test the networks level of generalisation. This is done through the analysis of the accuracy achieved through...