interactive_emulator.1

.\" Manpage for interactive_emulator
.\" Contact cec24@phy.duke.edu to correct errors
.TH interactive_emulator 1 "24.07.2012" "1.0" "interactive_emulator man page"
.SH NAME
interactive_emulator \- train, sample and query a Gaussian Process emulator
.SH SYNOPSIS
 interactive_emulator estimate_thetas INPUT_MODEL_FILE MODEL_SNAPSHOT_FILE [OPTIONS]
 interactive_emulator interactive_mode MODEL_SNAPSHOT_FILE [OPTIONS]
 interactive_emulator print_thetas MODEL_SNAPSHOT_FILE
.SH DESCRIPTION
interactive_emulator is an all in one program for training and sampling a gaussian process emulator. The format of the input MODEL_SNAPSHOT_FILE for interactive_mode must match the format of the output MODEL_SNAPSHOT_FILE from estimate_thetas
.SH estimate_thetas INPUT_MODEL_FILE MODEL_SNAPSHOT_FILE [OPTIONS]
Estimate_thetas trains an emulator on the given INPUT_MODEL_FILE producing a MODEL_SNAPSHOT_FILE that can be used for sampling the emulator at a later date. 
.P 
Options which influence estimate_thetas are:
 --regression_order=0 (const)
 \--regression_order=1 (linear)
 \--regression_order=2 (quadratic)
 \--regression_order=3 (cubic)
 \--covariance_fn=0 (POWER_EXPONENTIAL)
 \--covariance_fn=1 (MATERN32)
 \--covariance_fn=2 (MATERN52)
 (-v | --pca_variance=) FRAC: sets the pca decomp to keep cpts up to variance fraction FRAC
.SH interactive_mode MODEL_SNAPSHOT_FILE [OPTIONS]
Options which influence interactive_mode:
 (-q | --quiet): run without any extraneous output
 (-z | --pca_output): emulator output is left in the pca space
.P
The trained emulator represented by MODEL_SNAPSHOT_FILE is sampled at locations given on STDIN, as:
.P
 X_sample[0] ... X_sample[number_params-1]
.P 
Normally the number_outputs dimensional emulator mean and variance are output on STDOUT as:
.P
 E(Y[X_sample,0])
 V(Y[X_sample,0])
 E(Y[X_sample,1])
 V(Y[X_sample,1])
 ...
 ...
 E(Y[X_sample,number_outputs-1])
 V(Y[X_sample,number_outputs-1])
.P
Running with with (-z | -- pca_output ) will instead produce output in the PCA space, this output will have dimension < number_outputs but it otherwise produced identically.
.SH print_thetas MODEL_SNAPSHOT_FILE
Output to STDOUT the length scales for the covariance function as estimated by estimate_thetas. The pca-transformed design + training data for each principle component emulator is written to pca_emu_summary_<index>.dat where index runs from 0,...,number_outputs-1.
.P 
The pca_emu_summary_<index>.dat files are laid out as
 X[0, 0] X[0,1] ... X[0, nparams-1] Y_pca_<index>[0]
 X[1, 0] X[1,1] ... X[1, nparams-1] Y_pca_<index>[1]
 ... 
 X[N-1, 0] X[N-1,1] ... X[N-1, nparams-1] Y_pca_<index>[N-1]
Where N = number_model_points in the above.
.SH INPUT_MODEL_FILE format
An input model is specified interms of the locations in the parameter (design) space where it was sampled and the values of the model produced at these locations. 
.I 
 number_outputs
The dimension of the output vector we wish to emulate, y_out = y_1...y_t
.I 
 number_params
The dimension of the parameter space we are moving around in 
.I 
 number_model_points
The number of samples of the model (y_out) we've taken through the number_params dimensional parameter_space.
.P
In the following example we label the design data as X and the training data as Y.
  BEGIN EXAMPLE
    number_outputs
    number_params 
    number_model_points
    X[0,0]
    ...
    X[0,number_params-1]
    X[1,0]
    ...
    X[number_model_points-1,number_params-1]
    Y[0, 0]
    Y[0, 1]
    ...
    Y[0, number_outputs-1]
    ...
    Y[1, 1]
    ... 
    Y[number_model_points-1, number_outputs-1]
   END EXAMPLE
.P
number_outputs, number_params and number_model_points should be positive integers.  X[i,j] and Y[i,j] will be read as double-precision floats.

.SH BUGS
Don't believe anything this tells you without a lot of testing.
.SH AUTHORS
H.Canary (hal@cs.unc.edu)
C.Coleman-Smith (cec24@phy.duke.edu)