The SGDQN paper has been published
on the JMLR site.
This variant of stochastic gradient got very good results
during the first PASCAL Large Scale Learning Challenge.
The paper gives a lot of explanation on the design of the algorithm.
Source code is available from Antoine's web site.