Differences

This shows you the differences between two versions of the page.

Link to this comparison view

goplayer:mcmpgmmdist [2012/07/11 15:34]
aiartificer [Scratchwork Notes]
goplayer:mcmpgmmdist [2012/07/13 07:38] (current)
aiartificer [Explanation of the Monte Carlo Multilayer Perceptron Gaussian Mixture Model Distribution (MCMPGMMDist).] Describing drawbacks of this approach
Line 13: Line 13:
 Below are comparative estimates of the performance and memory footprint of this approach versus using a frequency table: Below are comparative estimates of the performance and memory footprint of this approach versus using a frequency table:
  
-|                    ^ Read Performance ^ Memory Footprint   ^+|                    ^ Read Performance ^ Memory Footprint              ^
 ^ Frequency Table    | O(1)             | Polynomial and/or Exponential | ^ Frequency Table    | O(1)             | Polynomial and/or Exponential |
 ^ Percpetron_calc    | O(1)             | 4k + 4                        | ^ Percpetron_calc    | O(1)             | 4k + 4                        |
 ^ GMM_sample         | O(1)             | 4k + 77                       | ^ GMM_sample         | O(1)             | 4k + 77                       |
  
 +There are three drawbacks for this approach.  First, the mechanics of the relationship between the conditionals to produce the given node cannot be easily decomposed to component parts when the node is recursively expanded to a MEBN.  Second, a Gaussian mixture model might provide more expressive power for the distribution than needed, leading to learning inefficiencies.  A four parameter beta distribution might be better.  Finally, observations cannot be explicitly added to the distribution as can be done with a frequency table.
 ===== OpenCL Code for MCMPGMMDist ===== ===== OpenCL Code for MCMPGMMDist =====
  

goplayer/mcmpgmmdist.txt · Last modified: 2012/07/13 07:38 by aiartificer