I have two points.
First, given the growing popularity of models where distributions make the state, Python resources for (adaptive) sparse grids:
—http://onlinelibrary.wiley.com/doi/10.3982/ECTA12216/full
—http://garcke.ins.uni-bonn.de/research/pub/sg_for_rl.pdf
would very be useful. At the moment, we have http://sgpp.sparsegrids.org/index.html. There is also TASMANIAN (not from Tasmania): http://tasmanian.ornl.gov/documents/UserManual-1.pdf.
Second, what are people’s thoughts on the various potential methods to do high dimension approximate dynamic programming? Adaptive sparse grids seem ideal, but require a bit of engineering. Another option seems to be neural nets (Bertsekas, Neureudynamic Programming), but I am not sure how one reduces the set of grid points. Both seem to suffer from convergence issues.