[Work Log] FIRE = Cluster model
May 08, 2014
Implemented cluster model in cluster_model.{cpp,h}. Implemented synthetic data generation in synthetic.{cpp,h}. Working on initial model estimation using k-means.
Run #1: Adding estimation of epsilon.
Description: After estimating 'A', also estimate the noise scale, epsilon (assumed 1.0 until now).
Method: Project data onto 'A' in data space, take mean projection error (see kjb_c::project_rows_onto_basis).
Results:
Using known epsilon: 0.5
-------------------
Training error: -45.3955
Prediction error: -46.5
Estimating epsilon: 0.46192
----------------------
Training error: -45.7958
Prediction error: -47.1319
Discussion: in the ballpark!
Run #2: initial cluster estimation
Description: Iteratively estimate three clusters and assign memebership.
Issues: third cluster initialization is identical to second. Guess: it's picking the same bad point over and over.
Run #3: initial cluster estimation, per-cluster obs. model
Description: Like Run #2, but each cluster has an individual observation model
Posted by Kyle Simek