r - Neural network inputs transformation -

im trying implement neural network following data. have decided transform output variable 0-1 , use logistic activation function. question how transform input variables? have read commonly transform them either 0/1 or -1/1 , use either logistic or tanh hidden notes transformation function. als read transform input data dummies ie. (1,2,3) or not transform @ all. im kind of puzzled optimal do. second question concerning matter how choice hidden nodes activation function, based on data transfomrmation?

thank s.

      age   salary mortrate clientrate savrate partialprate  [1,]  62 2381.140    0.047       7.05     3.1            0  [2,]  52 1777.970    0.047       6.10     3.1            0  [3,]  53 2701.210    0.047       6.40     3.1            0  [4,]  52 4039.460    0.047       7.00     3.1            0  [5,]  56  602.240    0.047       6.20     3.1            0  [6,]  43 2951.090    0.047       6.80     3.1            0  [7,]  49 4648.860    0.047       7.50     3.1            0  [8,]  44 3304.110    0.047       7.10     3.1            0  [9,]  56 1300.000    0.047       6.10     3.1            0 [10,]  50 1761.440    0.047       6.95     3.1            0 [11,]  63 1365.660    0.047       6.40     3.1            0 [12,]  51  986.530    0.047       6.40     3.1            0 [13,]  81    0.000    0.047       8.10     3.1            0 [14,]  64    0.000    0.047       5.80     3.1          697 [15,]  73    0.000    0.047       6.90     3.1          197 [16,]  56  226.890    0.047       5.15     3.1          750 [17,]  51 2576.645    0.047       3.70     3.1         8207 [18,]  66 3246.710    0.047       4.30     3.1            4 [19,]  66 3105.950    0.047       4.50     3.1         2998 [20,]  64  114.950    0.047       6.60     3.1          500 [21,]  84 1468.030    0.047       4.30     3.1         5000 [22,]  55 2616.510    0.047       4.70     3.1         3629 [23,]  71 3189.680    0.047       5.90     3.1         5445 

you can refer great , vast neural network knowledge repository:

ftp://ftp.sas.com/pub/neural/faq.html

normalization theoretically not needed, can achieved other values of input weights. in practice idea - otherwise during learning stumble upon arithmetic overflow in neuron weights. range [0;1] better [-1;1] tricky question needs diving detailed analysis of data set, , learning algorithm.

please refer ftp://ftp.sas.com/pub/neural/faq2.html#a_std.