Very interesting thread.

In 2004's Deep Fritz vs. Kasporov chess match, Fritz = four 2.8 GHz Xeon processors. In this year's match vs. Kramnik, Fritz was purportedly given 'much more' processing power.

If it takes 4+ super-duper processors to figure out the best next move for a single chess player... and we're asking a program/game to figure out the best next move for multiple players/factions with the processing power of our typical home computers, and to do it in less than 60 seconds (compared to our human taking 5-30 minutes to think)... we shouldn't be extremely surprised that some of those moves are gonna seem like blunders to us human players - one's we wouldn't have made.

I think the TW-series is more nuanced (or tries to be) than smash-and-bash, first-one-with-the-most-provinces-WINS! race-type playing style. But I grant the point that on VH/VH settings, the AI almost forces that.

Hence I conclude that the optimal setting for a full, deep, using-every-asset available (not just military) game, where the AI is fully engaged, is: Medium.