Fullscreen
Loading...
 
Print

General recommendations

Experiments and validation

It is a good idea to start with a Goal of Experiments section.
Your experiments want to demonstrate something;
the idea structure is to have questions.
Then, you plan one experiment to answer each question.
In the result part, you interpret the results of each experiment as answering the question.
This way, the reviewer will have a very clear and neat idea of where you are going and why.
Minimize the cognitive efforts, which is always an excellent idea.

Having a baseline

You must compare your stuff to something which is at the state of the art and is not your work (unless you are the pope of the domain of course).
You must tune the hyper-parameter of the competitor algorithm in a fair way.

Sensitivity analysis

Your algorithm has hyper-parameters. ALWAYS do a sensitivity analysis. In some cases, this is a complementary result; in any case, it feels more comfortable than saying I chose the value of .35 after preliminary experiments.

About doing new experiments

If you expect these experiments to take long, it's a good idea to ask your PhD advisor's opinion before. If not, just try.

Note that a new experiment should have a goal: you want to show this or that.

Note that you must not phrase it as: I want to do this experiment to show that my algo is better than XX.
When you do an experiment, you have of course a preference about the results; BUT YOU MUST APPEAR TO BE FAIR. NOT FAVORING YOUR ALGORITHM.

Before discussing with your PhD advisor about experiments

  • Google the literature
  • Make a list of problems which are relevant to your goal
  • See who has worked on each of those problems
  • Rank them according to what these problem allow to show (not judicious to select very similar problems)

Before showing results to your PhD advisor

Check.


Contributors to this page: sebag .
Page last modified on Monday 11 of March, 2013 19:02:29 CET by sebag.