Problem of ANN , Not predicting results other than trained data properly ?

(a) This is a typical 2 input 1 output problem . (b) I have normalized the input data between 0.2 - 0.8 (c) I have developed a ANN network (nnet), and used those weights which gives the least error. (d) The predicted results are excellent (for the training + testing + validation) (e) but when I use the network for a new set of data , say I give the two inputs as 0.9 0.9 , it gives absurd results and does not even follow the trend i.e. as the two inputs increase the output should increase.

Any help would be highly appreciated. Thanks,