40m QIL Cryo_Lab CTN SUS_Lab TCS_Lab OMC_Lab CRIME_Lab FEA ENG_Labs OptContFac Mariner WBEEShop
  40m Log  Not logged in ELOG logo
Entry  Thu Sep 20 08:50:14 2012, Masha, Update, MachineLearning, Machine Learning Update 
    Reply  Thu Sep 20 22:52:38 2012, Den, Update, MachineLearning, Feedback controller 
       Reply  Fri Nov 2 13:20:35 2012, Masha, Update, MachineLearning, Feedback controller standard_BATCH_0p35_ref_plant_lc.pngstandard_QUICKPROP_0p35_ref_plant_lc.pngstandard_RPROP_0p35_ref_plant_lc.pngstandard_INCREMENTAL_0p35_ref_plant_lc.pngstandard_INCREMENTAL_0p9_0p9_ref_plant_lc.png
Message ID: 7418     Entry time: Thu Sep 20 08:50:14 2012     Reply to this: 7424
Author: Masha 
Type: Update 
Category: MachineLearning 
Subject: Machine Learning Update 

 Hi everyone,

I've been working a bit on neural network code for a controller, and thus far I have code that creates a reference plant neural network. This is necessary to perform a gradient-descent learning algorithm on a controller neural network (one that reads an error signal and outputs actuation force). Because the error signal is read after the previous output of the controller neural network has passed through the plant, in order to calculate the gradient, either the inverse of the plant needs to be calculated, or the plant can be simulated through a neural network, and the error signal can thus be back-propagated through the plant neural network to find the gradient with respect to the output (as opposed to with respect to the plant), and thus back-propagated through the controller network in order to learn. 

I have uploaded to my directory a directory neural_plant. The most important file is reference_plant.c, which compiles with the command

 gcc reference_plant.c -o reference_plant -lfann -lm

The code runs on a file called reference_plant.data, which consists of a series of delayed inputs i_1, i_2, i_3 ... i_{n - 1} of the plant signals and then an output that is i_{n}, the subsequent signal. Parallel streams may also be used, if more than one signal is to be read. The top of the file must contain the number of total training packets (input-output groups), followed by the number of inputs, and the number of outputs. reference_plant.c also has constant variables which specify the number of hidden neurons, which can be changed. 

All of this code runs on the FANN library. If the code doesn't seem to be compiling, then it means the library might have to be downloaded and built from source.
Thus far, I have created my own plant in simulink (the driven, damped harmonic oscillator, as before), and obtained results of 0.0002929292 training MSE after 5 epochs (subsequently lowered to 0.000)  and 0.000 training error. This, however, is due to the fact that my plant is overly simple, and seems only to need 3 time-delayed plant signals, rather 31 to specify it (since all motion is second-order). 
It should be fairly easy to use interferometer signals as input to this plant by just reading some signals and parsing them into time-delayed groups. (I tried this over the summer with my previous code, and it seemed to work, although I haven't accessed any of the channels to obtain data lately). 
In terms of LIGO stuff this week, I'm going to be finishing up (writing) my final report, but please let me know if you have any comments or concerns. 
ELOG V3.1.3-