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.