2.10 What is better to obtain from the network - number or a decision? (translation by Ćukasz Brodziak; lukasz.brodziak@gmail.com) While using a neural network one has to remember that results returned by the network even if they are numeric values (which can be subject to appropriate scaling) always are approximate. Quality of the approximation can vary, yet the accuracy of significant figures is out of the question, it is good if a product returned by a neuron has accuracy higher than two significant figures (which means that the fault can reach several percent). That is simply what the nature of the tool is. Awareness of the occurrence of such limitations is forcing an appropriate interpretation of the output signals, so that they can be used in sensible way, and induces thinking about which neural calculations model You want to use. Generally speaking, neural networks can form to types of models: regressive and classification model. Regressive model is a type of model in which, on the output of the network, we expect (and demand) specific numeric value, which is a solution of the problem. Interpretation of such model is presented on figure 2.34. Fig. 2.34. Example of neural network regressive model In the model shown on the above figure the task of the neural network is to estimate the price of the apartment. On the input we load data that can be numeric values (such as area in square meters) and data that are nominal values (for example if the apartment has a garage assigned to it), on the other hand on the output we expect a numeric value which would show the amount which can be gained on sale. As all apartment buyers and sellers know the market price of the apartment depends on many factors, yet nobody can present strict economic rules which would allow to predict that one apartment will cost this much, and the other one will cost twice as much. Seemingly it is impossible to predict because the price is each time a result of an unconstrained market game and sovereign decisions made by people who buy and sell a specific apartment. Still it appears that a neural network after long training (based on the previously finalized buy-sell transactions data) can form a regressive model of the problem so good that real prices obtained in successive transactions were only few percent different from the ones „predicted” by the network. The alternative classification model is connected with a demand of obtaining from the network information on classifying an object described by input data to one of the classes. Of course in this type of problem on the output of the network a nominal variable will be placed (see figure 2.35). Fig. 2.35 Example of neural network classification model. Example shown on the figure regards the problem of classification of the people (or companies) soliciting for a bank loan. Such client has to be classified by the bank worker to one of two categories: First one contains clients who are trustworthy and reliable, and therefore can get a loan because they are going to pay it back along with the interests and will give the bank profits. To the second category are numbered all the frauds and potential bankrupts who will not pay the loan exposing the bank to losses. How one can distinguish one category from the other? There are no strict rules or algorithms but correct answer can be provided by neural network taught with data from the past. There is usually a lot of such data as very bank has given a lot of loans and has full data of its clients both those who paid the loans back and those who did not. Years of experiences in neural network exploitations proved that it is most convenient to interpret the tasks given to a neurocomputer in a way that would allow the answer to be given using classification model. For example we can demand that the network will state whether profitability of the investment is „low”, „average” or „big” relatively whether the loan-taker is „reliable”, „risky” or „totally unreliable”. However the demand of specific estimation of the amount of profit, accuracy of the risk level or the limit of the amount that can be loaned to somebody will inevitably lead to frustration because in general it is the thing which network cannot do. That is why the number of the outputs of the created network is usually bigger than number of questions that we seek an answer to. This happens because for many output signals we have to artificially add few neurons servicing each output – for example in a way that would allow the predicted range of output signal values to be divided into certain sub-ranges which distinction is essential for the user. In such case we should not force the network to return specific „guessed” value on its output instead we work in a way that each output neuron is responsible for signalizing appurtenance of current answer to specific range and usually that is all we need. Such classification network is way easier to build and teach, while making a network from which we „squeeze out” specific answers to mathematical problems is typical „art for an art's sake” – time-consuming fun with minimal practical application. Above formulated conclusions are summed up by figure 2.36 which presents one of the „classic” problems presented to neural network: prediction of currency rate. Fig. 2.36. Currency rate estimation is one of many problems in solving which the creator of a neural network can choose between regressive and classification model. As shown on the figure the problem can be solved in two ways: Either one can build a prognostic model which attempts to determine how many Euros will a dollar be worth tomorrow or one can be satisfied with a model which only will signalize weather one should expect a raise or fall of the rate value on the next day. The second task will be easier for the network to solve and the prediction of rate value raise or its fall can be useful for someone wanting to buy dollars before travelling abroad.