-
It would be neat to have them as crew if possible. They could eat enemy NPC's while in combat too. It's potentially game-breaking - gotta warn about that - but it could be fun.
-
Relationship between the contributions and different training schemes (e.g. cutting-plane, sub-gradient, Frank-Wolfe).
-
It turns out that self loops and tottering is hurting predictions. SSVM does not take care of this. To avoid tottering, we need to solve an ILP.
The task is to update structured prediction learning t…
-
Focus on three methods: SSVM (single label), multi-label SSVM and RankSVM, baselines and other methods can be selected later.
-
Given the same weight vector for SSVM, does the prediction by the list Viterbi algorithm is the same as that of ILP (assuming no ties)?
Write down the proof.
-
Feature scaling is necessary because SVM doesn't work very well if all features are positive + sum of unary features should be approximately equals the sum of pairwise features.
-
Prove the inference problem in SSVM, i.e., find a maximum weighted simple path with exactly k edges in a complete directed graph given a source node, is NP-hard.
-
As the code in notebook ssvm.ipynb and ssvm_ml.ipynb is quite similar, refactor and merge them into one notebook.
-
4 variants of setting up: i.e., single-/multi-label SSVM and with/without sub-tour elimination during training.
-
Write a well documented file that has functions to be imported.