I propose to:
a) assess what blocklimit is currently technically possible without driving up costs of running a node up too much. Most systems currently running a fullnode probably have some capacity left.
b) set the determined blocklimit at the next reward halving
c) then double the blocksize limit at every halving thereafter up to a hardlimit of 8GB.
Doubling every four years will stay within expected technological growth. Cisco's VNI forecast predicts a 2.1-fold increase in global average fixed broadand speed from 2014 to 2019. Nielsen's law, which looks more at the power user (probably more fitting) is even more optimistic with +50% per year.
This proposal can be considered a compromise between Pieter's and Gavin's proposals. While the growth rate is more or less what Pieter proposes, it includes an initial increase to kickstart the growth. If we start with 8MB, which seems to be popular among miners, we would reach 8GB in 2056 (as opposed to 2036 in BIP101). The start date (ca. mid 2016) is also a compromise between Pieter's 01/2017 and Gavin's 01/2016.
It's simple, predictable and IMHO elegant -- block subsidy halves, blocksize limit doubles.
It might make sense to update the limit more often in between, though. Either completely linearly based on a block's timestamp like in BIP101, or for example for each difficulty period.