AIC: takes a "prediction" point of view
BIC: Asymptoticaly accurate for model selection in context of expected loss -- to minimize expected loss in "betting" (Bayesian) sense
BIC tends to pick smaller models than AIC (for n>8)
FYI:
Kadane JB, Lazar NA (2004). Methods and criteria for model selection. J Am Statist Assoc,99,
279-290.
Web link:
http://docserver.ingentaconnect.com/deliver/connect/asa/01621459/v99n465/s26.pdf?expires=1144761197&id=28169785&titleid=30000015&accname=Welch+Medical+Library&checksum=AEC1C45051FE5A977928C5C91333EC84