SQL Apprentice Question
I am working
on a project that is related to the field of agriculture and that has
as an objective to find the "optimal values" of the operating
conditions that affect the outcome (the amount of meat produced i.e.
the weight) of an animal production (chicken broilers in my case). To
do so, I have to use historical data of previous productions as my
training dataset. The length a production cycle is typically around 44
days. For each production, a data acquisition system stores the
real-time and historical data of hundreds of parameters. These
parameters represent sensor measurements of all the operating
conditions (current temperature, set point temperature, humidity,
static pressure, etc...) and these are what I refer to as the inputs.
The operating costs and the production outcome are what I refer to as
outputs. The operating cost is indirectly computed from parameters
like water consumption, feed consumption, heater/cooling runtimes, and
lighting runtime; and the outcome of a production is defined by
parameters like animal mortality and conversion factor (amount of feed
in Lbs to produce 1Lb of meat). So the main objective of this project
is to find the set of "optimal daily values" (1value/day) for the
inputs that would minimize the operating costs and conversion ratio
The biggest problem I am facing right now is the following: The
historical data that I have in the DB are time series for each measured
parameter. Some of these time series follow some kind of cyclic
pattern (e.g. daily water/feed consumption ...) while others follow an
increasing/decreasing trend (animal weight, total heater run time,
total water/feed consumption.....). My goal is to be able to come up
with a model that suggests a set of curves for the optimal daily values
throughout the length of the production cycle, one curve for each
measured input/output parameter. This model would allow the farmer to
closely monitor his production on a daily basis to make sure his
production parameters follow the "optimal curves" suggested by my
model. I have looked at ANN and I think it might be the solution to my
problem since it allows to model multiple input/outputs problems (Am I
wrong?), but I could not figure out a way to model the inputs/outputs
as time series (an array of values for each parameter). As far as I
know, all kinds of classifiers accept only single valued samples.
One approach would be to create one classifier/day (e.g. for day1:
extract a single value for each parameter and use these values as a
training sample and repeat this for all previous production to
construct the training set). The problem with this approach is that 44
or so classifiers will be constructed (hard to manage all of this) and
each of these resulting ANN will be some kind of "typical average"
of the training data but not necessarily the "optimal values"
leading to the best production outcome, if I am not mistaken.
Another approach would be to find a way to feed in the inputs and
outputs as time series (an array of 44 daily values for each
input/output parameter). In this case, there would be only one
resulting ANN and the training samples, would be a set of arrays for
each parameter, as opposed to single daily parameter values in the
first case. The problem is, I could not find any classifier that would
allow me to do that.
Another issue that I have is the amount of data. While a single
production cycle could represent 1-2GB of data, the length of the
production cycle (44 days) makes it difficult to have 100's of
production cycle historical data, as I could gather data for no more
than 7 full cycles/year. Fortunately, a farm can have many production
units (5-10 barns/site in big sites), so this makes it possible to have
40-70 cycles/yr. My question is: would this be enough to come up with
an acceptably accurate model or is it necessary to have hundreds of
I seem to remember "Evolutionary Operation" -- EvOp -- from chemical
manufacturing. The basic idea is small adjustments in multiple factors
to get an optimal setting for a process. There was an assumption of a
local optimal point among the parameters, but a relatively small sample
is needed to adjust things.