optimization question
Kartit
I am new to machine learning, and I am experimenting with neural nets.
I would like to know if the they are suitable for my problem and I am
lost between linear programming/genetic algorithms and neural nets. I
am trying to optimize the number of doctors to allocate in order to
minimize patients non treated in case of an emergency.
my data sets format as follows:
1. Event / Patients treated / patients non treated
2. Event / Doctor / speciality.
3. Event / number of doctors involved in best case / worst case
scenario
my constraints are: number of patients non treated should not exceed a
parameter X.
Output: How much doctors to allocate from each speciality.
I hope this explains my problem, I am just asking for a good approach
if this should use linear programming the finding an optimal solution
using genetic algorithms? any suggestions I would appreciate it.
Clif Davis
Not enough information.
Your data set format generates some questions about the meaning of the
data. More information is required to give specific advice. Does
event give the key to a date or relevant time period? Can Doctors
have multiple specialties? Are type 1 events broken down by
specialty? Presumably best case scenario means minimal number of
doctors required for exactly X untreated patients, but what does worse
case mean? Many more doctors needed? No patients treated? The
exact form of the problem is unclear.
If you know the probability distribution of number of patients for
each specialty and the cost of a doctor in each specialty and the cost
of an untreated patient in each specialty, then it's just a problem in
statistics. If you have a simple probability distribution of the
number of patients for each specialty and the requirement that the
number of patients not exceed X with some probability then you have a
linear programming problem. If you only know the number of treated
and untreated patients each day with no knowledge of the specialty
required, you have a harder problem, particularly if the number of
doctors in a best case/worst case scenario is also not broken down by
specialty.
The data set might provide historical test cases to see how well the
output of model does against the historical record, ie to see if the
best case would be achieved by the model. But what is the input of
the model other than X, the maximum number of allowable untreated
patients? Presumably we would have input with some clues as to the
anticipated patient load by specialty. Perhaps the input would be
number of patients by specialty from the previous day if doctors are
allocated daily?
It is possible to give some general advice without recourse to
details. If having an optimal answer is important and a linear
programming approach is feasible for your problem and computer
resources, then by all means use a linear programming approach. If a
linear programming approach or straightforward statistical analysis is
infeasible for some reason then the order in which I would try machine
learning approaches to this type of application are:
(1) Linear Regression tree [basically a decision tree with a linear
model at each leaf]. (and I might try a simple Regression tree with a
single value at each leaf first). [CART will now produce Regression
trees as I recall and SECRET comes to mind as one algorithm for Linear
Regression trees, but don't hold me to it.)
(2) Genetic Algorithm (but only if I knew the general form of what a
solution model would look like)
(2) Neural Network
(3) Genetic Programming (but only if I didn't know what the general
form of a solution model would look like and I had a LOT of CPU cycles
to throw at the problem.)
Clif Davis
And we see I can't count.
Please correct to:
(1) Linear Regression Tree
(2) Genetic Algorithm
(3) Neural Network
(4)Genetic Programming