Database Reference
In-Depth Information
In particular, we have pointed out a possible way to circumvent the complexity
issues related to the more sophisticated yet more realistic model of a
k-
MDP. A case
for this particular approach has been made by showing it to be consistent with the
classical approach of assuming a 1-MDP. Yet, it is too early to assess the actual
practical value of the approach, and there are still a lot of difficulties as regards
implementation to overcome. Mathematically, we still lack deep results on approx-
imation quality as well as numerical experience. With regard to recommendation
engineering, we still need to confirm that assuming a
k
-MDP rather than a classical
MDP actually results in a significant improvement and, if so, how to choose the
order
k
, which trades off between accuracy and simplicity. Furthermore, we need to
ensure that the above-presented factorization approach provides a satisfactory
approximation and figure out a way to choose the underlying partition suitably, for
example, by virtue of the above-outlined genetic algorithm framework, the practical
effectiveness of which, too, still needs to be evaluated.
Further, we have described how the approaches presented in the previous
chapters are related to our generalized
k-
MDP formulation.
With regard to future research, apart from the reasoning presented in this
chapter, it might be interesting to state the two approaches, i.e., the factorization-
based and the control-theoretic, within a unified framework. Such a framework may
then enable to discover more possible ways to connect the two approaches.
