Why does the discriminator loss converge but the generator loss not converge for a traditional GAN?

[wangyongjie-ntu]

For the earliest GAN implementation, the discriminator loss decrease but the generator loss increase. And the summed loss is increasing. Can anybody face these issues?

Another question, for a well-trained GAN. Can the discriminator return the probability of a training point? I.e., use the discriminator to detect the outliers.