problem of choosing right statistical method for scheduler prediction

Question

I am struggling to choose a right data prediction method for the following problem. Essentially I am trying to model a scheduler operation, trying to predict its scheduling without knowing the scheduling mechanism and having incomplete data.

(1) There are M available resource blocks that can carry data, N data channels that must be scheduled every time instance i

(2) Inputs into the scheduler:

• Matrix $X_i$ size M by N, consisting of N column vectors from each data source. Each of M elements is index from 1 to 32 carrying information about quality of data channel for particular resource block. 1 - really bad quality, 32 - excellent quality.

• Data which contains type of data to be carried (voice/internet etc)

Scheduler prioritizes number of resource blocks occupied by each channel every time instant i.

Given that

• I CAN see resource allocation map every time instant

• I DO have access to matrix $X_i$

• I DON'T know the algorithm of scheduler and

• I dont have access to the type of data to be scheduled.

I want to have a best guess (prediction) how the data will be scheduled based on this incomplete information i.e, which resource block will be occupied by which data channel. What is the best choice of prediction/modelling algorithm? Any help appreciated!

Answer 1

Do you know if the scheduler has a memory?

Let us assume for a moment that the scheduler has no memory. This is a straightforward classification (supervised learning) problem: the inputs are X, the outputs are the schedules (N->M maps). Actually, if every N gets scheduled and the only question is which M it gets, the outputs are lists which channel (or none) is scheduled to each block, and there is only a certain possible number of those, so you can model them as discrete outputs (classes) with their own probabilities. Use whatever you like (AdaBoost, Naive Bayes, RBF SVM, Random Forest...) as a classifier. I think you will quickly learn about the general behavior of the scheduler.

If the scheduler has a memory, then things get complicated. I think you might approach that as a hidden Markov model: but the number of individual states may be quite large, and so it may be essentially impossible to build a complete map of transition probabilities.