Database Reference
In-Depth Information
6.4 Summary
As we have seen so far, hierarchical methods are extremely important in many
scientific areas. We have described that especially multilevel methods, which
originate from numerical analysis, are in principle well suited to speed up the
reinforcement learning. However, their application to RL is quite difficult. We
used some specifics of recommendation engines to make the multilevel approach
more applicable in this field. However, more research is required here.
We now come to the outlook for the future. We have the choice between the use
of predefined hierarchies (as in the analytical case) or automatically generated
hierarchies (coarsening, as in the algebraic case). Currently we are working with
predefined hierarchies. Now these are hierarchies such as shop taxonomies or
product groups specified by the shop or category manager and not primarily
intended for the use in hierarchical preconditioners. Thus, they do not prove optimal
for this. (In practice, they are usually subjected to a comprehensive preprocessing.)
However, classic coarsening is also not the best possible method either, since of
course the existing category information should be exploited for the hierarchies. It
is therefore useful to employ a combination of the coarsening procedure, having
regard to the product attributes and hierarchies. This work is in progress.
