There's no way to do this automatically, but in your case just using the default 'OutputNetwork' setting ( last-iteration ) would do it.
I'm not an expert here but it seems to me that the only way for your loss (which is what you are actually trying to minimize) to be rising while your accuracy is improving is if your network's confidence in the chosen output class is getting worse (but it's still choosing the right one). It may be that you don't really want that network therefore. However, the number of images in your validation set may be a factor here. Perhaps you should increase the size of the validation set?
