Skal også sammenligne med MAUT.

Studerer nevrale nett for å sammenlikne med DRSA/DOMLEM.

Prøver å finne en god kilde som beskriver nevrale nett.

An arti FIcial neural network (ANN), or simply neural network (NN), is a

computational model of how a biological neural network works. To model the

biological neural network a set of neurons are de ned as numerical values.

A set of numeric weights represents the connections between the neurons.

These weights can be adjusted by a training method, such that the neural

network can learn simple concepts.

ArtiFI cial neural networks has been used in several real life applications

like image data processing, tra c control, regression analysis, pattern recognition

and classi cation like hand written character recognition.

Some studies have also tried to utilize neural networks in gene prediction.

According to Reese [18] the FI rst attempt was done by Brunak et al.[8]. Later,

similar methods have been in other works (References [18, 14, 12, 6]).

Scope of Work, Hypothesis line under work

Given the paper written last semester, and given we have the use cases, we can subtract information.

Artificial Neural Networks

NN technology, used for development of artificial intelligence. It is inspired by how the brains of humans

and animals work. The brain is composed of millions of neurons, and these neurons are

connected to each other by axioms and dendrites. The connections are adaptive, and so the

connection structure is dynamically changing. Changes of the connections is what we call

learning.

An artificial neural network is similar, where the strength of couplings between artificial neurons are described by numerical weights. One of the most used structure is the multilayer neural network. The neurons are usually modelled in three layers, where the first layer is called the input layer, the intermediate layer is called the hidden layer, and the last layer is called output layer. The neurons could also be referred to as nodes. The nodes of each layer are connected each other with adjustable weights. The process of adjusting these weights is called training. The numerical value of a unit is the sum of all its input connections times the weight of that input connection. The numerical value of each neuron is the adjusted with a non-linear activation function.

An ANN can be trained to recognize input patterns, give the desired result, by adjusting the weights. A neural network is therefore a general approximation function. A neural network approximation function can have an arbitrary number of inputs and outputs.

Feedforward calculation

After the weights have been trained in the neural network, it can be used to evaluate an input feature vector an predict an associated outcome vector. The process of taking an input vector and predicting an output is called feedforward operation. The feedforward operation is given by equations 2.1, 2.2, 2.3 and 2.4. The input vector is denoted by x.

Backpropagation

Found this, so far good, tutorial: http://www.ai-junkie.com/ann/evolved/nnt4.html

One negative, it is more complicated to construct the network:

Now we have defined our inputs and our outputs what about the hidden layer/s? How do we decide how many layers we should have and how many neurons we should have in each layer? Well, this is a matter of guesswork and something you will develop a ‘feel’ for. There is no known rule of thumb although plenty of researchers have tried to come up with one. By default the simulation uses one hidden layer that contains six neurons although please spend some time experimenting with different numbers to see what effect they may have. I’d like to emphasise here that the more you play around with all the parameters the better the ‘feel’ you are going to develop and the better your neural networks will be.