I have thought that what if different individuals had somewhat different weights for actions, representing different personalities. Brave creatures would be less likely to retreat, lazy ones less likely to work when tired, etc. There is probably some kind of mechanic like this already (tags?), but the real beauty of the system would come from these being inherited and/or learned. Dwarves could fall in love with other dwarves whose personality matched theirs closely enough and their babies could inherit the parents' "personality", for example by using an average of the parents' weights with some random noise added. Thus if a father was really talented miner, the child would probably be too. Of course really crappy dwarves (measured somehow) might not be able to find any partner at all, resulting in their removal from the gene-pool. Using only evolution would be too slow to produce (say) actual upper and working class of dwarves, however.
I'm not sure how the weight learning should happen for live dwarves, though, because the value of current behavior is difficult to measure as it depends much on circumstances and other dwarves. As a simple example, when facing a seriously dangerous assault, it might be best for few super-brave patriotic dwarves to stay outside and slow the enemy while the others ran inside and readied the defenses. If all ran, the whole fortress might get killed, but if all stayed, it would result in an unnecessary amount of deaths, so the optimal combination for the fortress in this situations would be to have a couple of fearless dwarves with the rest being cowards enough to stay alive.
I'll post more about this when I have the time.