How does a neural network work? Describe the structure and function of a neural network
(textual and, if necessary, with the help of an illustration).
• A neural network consists of multiple layers: it has an input layer, at least one hidden layer and an output layer, which each are linked.
• The artificial neural network is based on the biological neural network, where the input layer is represented by neurons that are linked with the dendrites of other neurons in the hidden layer. These neurons in the hidden layer are linked with neurons of another hidden layer or, if there is only one hidden layer, with the output layer.
• In this neural network information is given from one neuron to another in form of impulses.
• The neurons activate the given input on their synapses and form a weighted sum of the information (activation function, propagation function). This weighted sum information is given as an output to the next neuron as an input.
• The process of the transfer of information is not known, hence it is a black box model.
What requirements must be met for a neural network to be established?
• In order that a neural network can work, two special phases are obligatory: the training phase and the working phase.
• In the training phase the desired behavior is thought to the network. Therefore, you need sample data sets where the input and output constellations are known.
• When the network have “learned” this data, the working phase begins, which means that based on the training results, the neural network produces similar results for input values that correspond to the training sets.
• This implies that you need a high amount of sample data sets that are similar to the ones from the working phase in order to “train” your neural network properly.
• High amount of data sets (>100000)
• Both phases must have similar data sets
Specify the main differences between a fuzzy logic model and a neural network.
Fuzzy logic
Neutral Network
Influencing factors are known and include the relevant information
Influencing factors are not precisely known but sample data is available
Number of influencing factors is small (<20)
High-dimensional input space
simple model design
based on learning, trains itself from learing data sets
Typical white box model
black box model
helps to recognize patterns and behavior
helps to predict
Combination of Neuro and Fuzzy:
Fuzzy logic systems can be used for knowledge representation in neural networks
• The fuzzification, computed for independent input variables, can be modeled in the input layer
• The fuzzy rules which compute the conclusion independently of one another, can be represented by the hidden layers. The rules can also be described in multiple layers
• The defuzzification of the variables is also carried out independently in the output layer
For a water management system, known input variables are available for the resulting
damage caused by floods and the causative influencing variables (precipitation, buildings,
etc.).
Which model type would you envisage to depict this real word system (neural network,
fuzzy logic model, combined neuro-fuzzy model)?
Justify your choice.
Fuzzy logic:
Represents house damages very well in a descriptive and classified manner
Neural networks:
Shouldn’t be used
• No linguistic data can be processed
• System cannot be modified → black box
• Not enough data sets available for neural networks
Possible, if enough data sets are given
Designate in each case a water management task for which you would like to use a neural
network or a fuzzy logic model.
Explain why you made the choice by using the model properties.
EXTRA
Would you choose fuzzy models or neural networks for the implementation of the Water Framework Directive? Describe the differences of both.
The Water Framework Directive tries to improve the state of water systems. In order to do so, the patterns and behaviors need to be understood. That is why I would use a fuzzy model. Furthermore, the water systems and their characteristsics (for example, the water quality) is very difficult to describe in sharp numbers, rule-based
models can be a better choice. In addition, relationships in ecological systems like water systems are very complex – that is why a fuzzy logic is mostly appropriate, since imprecisions und uncertainties need to be considered.
How to determine the maximum possible number of rules?
• Multiplication of the number of the classes of the respective input variables
• E.g. 2 input variables with each 3 classes result in a maximum of 3x3 = 9 rules or in general ic with c = number of classes & I = input varibles
Zuletzt geändertvor 3 Monaten
