INFERRING PARAMETERS FOR A FIXED NUMBER OF SOURCES.
Given measurements of concentrations are various locations, we can try
to infer the locations and intensities of the sources, as well as any
other unknown parameters, such as the noise level for the detectors.
Here, we'll consider such inference when the number of sources is
fixed to a known value.
To begin, we need to specify the structure of the model and the priors
for the unknown model parameters. This is done with specification
commands similar to those described in the example of generation of
data (Ex-src-g.doc). However, here we'll assume that the noise level
for measurements is unknown, with the log of the noise standard
deviation being uniformly distributed between log(0.01) and log(1).
Otherwise, the specifications are the same as were used when
generating the data. The specification commands needed are as
follows:
> src-spec logc2 3 0:5 / -10:10 -1:1 0:1
> det-spec logc2 0.01:1
> flow-spec logc2 test 1 0.08 0.0001 0.06 0.00015
We use a data-spec command that specifies the locations of the
detectors and their measured concentration values:
> data-spec logc2 3 1 / grid2 data-grid2-0.1-1
There are 3 inputs (detector location coordinates) and 1 target value
(detector measurement). The detector locations are on the grid of 126
points with coordinates in the file 'grid2', and the measurements at
these locations are in the data file 'data-grid2-0.1-1', which was
generated as discussed in Ex-src-g.doc.
Next, we specify the Markov chain Monte Carlo operations that will be
used to sample from the posterior distribution of the unknown values,
which for this example are the noise standard deviation and the source
intensities and locations. A wide variety of general-purpose
operations are possible, as described in mc-spec.doc (though some
operations are not allowed at present, because gradient computations
are not implemented). Also, some special operations for source
location models have been implemented, as described in src-mc.doc.
For this example, simple operations such as single-variable slice
sampling are sufficient. We can specify that each iteration does 50
sets of slice sampling updates as follows:
> mc-spec logc2 repeat 50 slice-1 1 1 D slice-1 1 1 S end
The "repeat 50 ... end" construct says that the operations in the
middle are repeated 50 times, with the results storing in the log file
only after all 50 repetitions. Not storing every possible state saves
on disk space.
The operations repeated here are single-variable slice sampling
updates of detector parameters (indicated by "D"), which for this
example is just the noise standard deviation, and single-variable
slice sampling updates of source parameters (indicated by "S"), which
here are the intensities and locations of the sources. The two
numbers after the "slice-1" specifications are the width of the
initial interval (here 1) and the maximum width of the slice interval,
in units of this initial width (here also 1, which means that the
original interval is not expanded). The specified interval width is
multiplied by the default stepsize for each parameter, as described in
src-mc.doc.
We can also specify an initial state for the Markov chain, otherwise a
default initial state will be used (see src-initial.doc). For this
example, the following command is used to set the initial noise
standard deviation to its maximum value of 1, letting other values
default:
> src-initial logc2 / 1
Using a large noise standard deviation initially may help the Markov
chain move around at the beginning, when the fit to the data is poor.
(With the slice-1 operations used here, however, this explicit
initialization to a large value may be unnecessary.)
Finally, we can actual run the Markov chain that samples from the
posterior distribution. The following command does 1000 iterations,
taking about 45 seconds on the system used (see Ex-test-system.doc):
> src-mc logc2 1000 &
The "&" runs the command in the background.
We can monitor the progress of the run by displaying the state for the
last iteration saved in the log file. For instance, after a few
seconds, we might see the following:
> src-display logc2
PARAMETERS OF SOURCE LOCATION MODEL AT ITERATION 341
Noise standard deviation: 0.1007
Wind speed: 1.0000
Q x y z
1 1.83 9.02 0.10 0.45
2 0.81 -6.33 0.71 0.88
3 0.55 4.42 -0.41 0.20
Here, "Q" is the source intensity. A specific iteration can be
displayed by specifying the iteration index as an additional argument
(see src-display.doc for details).
We can also plot traces such as the following:
> src-plt t n logc2 | plot
This plots the iteration number (t) on the horizontal axis and the
noise standard deviation (n) on the vertical axis, for all iterations
stored so far in logc2. It takes about 20 iterations for the noise
standard deviation to fall to the vicinity of the 0.1, which was the
value used to generate the data. We can also monitor the source
locations with a command such as
> src-plt t x@ logc2 | plot-points
Here, "x@" represents the x coordinates of all the sources. (The
"plot-points" program should plot multiple sets of data pairs as
points.) After about 50 iterations, the three sources have x
coordinates in the vicinity of the correct values of 4.5, -6, and 9.
Other quantities that can be plotted are described in
src-quantities.doc (see also mc-quantities.doc and quantities.doc).
It appears that 1000 iterations are more than enough. If more
iterations were needed, more could be done. For example, the command
> src-mc logc2 3000 &
would do an additional 2000 iterations (after which the total number
would be 3000). Commands such as this can also be used to continue
runs after a src-mc command was interrupted for some reason.
Once the run is complete, we can make predictions for the
concentration at any set of locations, using the src-pred program (see
src-pred.doc). Predictions at all locations in the file 'grid1' can
be made with the following command:
> src-pred inb logc2 100: / grid1 >predc2
This command writes the predictions to standard output, which is here
re-directed to the file 'predc2'. The characters in the first
argument, "inb", specify that inputs (i) are included in the output
(always four inputs, for x, y, z, and t, even if the data file had
fewer), that the mean (n) of the posterior predictive distribution
will be written, and that the output will be bare (b), with no
headers.
These predictions can be compared to the true values using the R
functions in data-plot.r. There is a good correspondence, except that
various spurious values are present at locations with x coordinates
less than -8. This happens because there are no measurements at
locations with x coordinates less than that (ie, downwind), so there
is no evidence that such sources don't exist.
We can also use the results of the MCMC run to find the expected total
intensity of sources in a set of grid cells. Here some example output:
> src-intensity 5 4 / logc2 100:
0.0000 -8.0000 -0.7500
0.0000 -8.0000 -0.2500
0.0000 -8.0000 +0.2500
1.4688 -8.0000 +0.7500
0.0000 -4.0000 -0.7500
0.0000 -4.0000 -0.2500
0.0000 -4.0000 +0.2500
0.0788 -4.0000 +0.7500
0.0000 +0.0000 -0.7500
0.0000 +0.0000 -0.2500
0.0000 +0.0000 +0.2500
0.0000 +0.0000 +0.7500
0.0000 +4.0000 -0.7500
0.5002 +4.0000 -0.2500
0.0000 +4.0000 +0.2500
0.0000 +4.0000 +0.7500
0.0000 +8.0000 -0.7500
0.0000 +8.0000 -0.2500
1.8144 +8.0000 +0.2500
0.0000 +8.0000 +0.7500
The src-intensity command asks for a grid over x and y, with 5 cells
in the x dimension and 4 in the y dimension. Iterations from 100
onwards in log file logc2 are used to estimated the expected total
intensity in each of these cells. The output has one line for each of
the 20 cells, with the first number being the estimated total
intensity, and the second and third numbers being the x and y
coordinates of the centre of the cell. Grids can extend over z also,
and over t as well for non-steady-state models. The output can be
plotted in R using the functions in data-plot.r.