I did a 12k book learn about 2 weeks into the age, having invested something like 170k books already. Wasn't too bad, since it was mostly in categories I was neglecting.
Also note that if you are waved, you'll usually want your armies at home in order to retaliate immediately. Sometimes that's worth having elites in and dealing with the losses as they come.
This is why having a lot of hospitals during normal and aggressive stance is a big deal, due to oow losses being much higher. Hitting an undead with armies in and no hospitals = very good loss of their offense; and if they have too much offense and too little defense, a 2x in that spot does huge damage.
If you were operating with a super-offense strategy, as a lot of top kingdoms do when rebuilding, then you will want to be active in attacking of course - but that does not consist of attacking blindly for crap gains. If you have high offense, you should have an intel machine that allows you to best utilize that said offense, and not hit anywhere just to say you hit something.
http://www.godlygaming.org/2011/04/1...ive-randoming/ - Godly's guide to effective randoming is a good read, once you subtract the obligatory trash talking. He'd probably call you total garbage for using a Learn attack, and 99.99% of the time he'd be right.
Also note that once you reach the point of diminishing returns on science, learns aren't all that great. It looks impressive on paper to say you have 4000 bpa or something, but it's usually not worth a Learn attack when you have ample science for your size already. Learn is a good way to build up your science if you're behind, and especially good if a target has high investment in areas where you are lacking.
Also note - gains science is awesome for learns. I started mid-age and put almost everything into gains science, and let the server fund about half the alchemy+housing science I had plus some others.
Finally - when waving a kingdom, enemy defensive losses matter, but for various reasons you might not want to gain acres. Learn is a decent alternative with faster turnaround time than trad march, and there isn't a good opportunity to use learn attacks in war. Even then, you wouldn't want to use your attack on a sucky learn, but the range of acceptable Learns widens during a wave. The 1.5 bpa/hr benchmark I would consider a good learn was for hostile waves, but if I'm randoming for Learns I'd want at least 2 bpa/hr, and preferably 3 bpa/hr.