Skip to content
Blocks Relocation Problem
C++ C Makefile Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
results17
src
Makefile
README
auto.sh
data.tar

README

Author   : Marco Caserta
Email    : marco dot caserta at ie dot edu
Revision : April 28, 2017


This part of the project is associated to the review carried out in April 2017.
Due to some tests, SV and Silvia pointed out a potential mistake in the CM
algorithm. The problem was that the maximum height (H+2) was not always
respected. Actually, the corridor method, in selecting a stack, was not
checking whether the maximum height of that stake was respected and, therefore,
it could end up piling up elements above the H+2 height limit. 

I believe that, in the original OR Spectrum paper, this constraint was not 
enforced. I think that, while this is common practice now, it became an accepted 
practice for this instances from the EJOR paper onward. In any case, the code
has been fixed in April 2017 (a comment has been added within the code, to
indicate the change), and the whole set of experiments have been re-ran.

I also observed another issue related to the H+2 limit and the corridor. Assume
we have a corridor of 5 stacks in the current bay. However, only 3 of the
stacks have a current occupation of less than H+2. In other words, while we
would like to identify 5 candidate stacks, only 3 of them have not reached yet
the maximum height. In this situation, we cannot build a corridor of width 5.
We will have to do it with a corridor of size 3. Thus, the actual corridor is
now built as the minimum between the required corridor size and the number of
still usable stacks (stacks that have not reached yet the maximum height).


Analysis of the Results
-----------------------

The file "resultsAnalysis.Rmd" has been used to analyze the raw data, clean up
the files, and compute the average values. The output is the table
"avgTable.txt", which provides the average number of moves for each instance
class (averaged over 40 instances per class.)


Results
-------

- The directory "results17" contains the summary of the results, i.e., the
  recomputation of the tables of the OR Spectrum paper. The file avgTable.txt
  provides the average number of moves for each instance class. In addition, we
  verify that, for each instance class, 40 problems have been solved.
- Within that folder, there is another folder, called "rawData", which contains 
  the detailed results for each instance within each class. 
- A few output files have also been created, i.e., output5 and outputLarge,
  with the output of the algorithm. I used these files to check that the
  solutions found by CM do not exceed the limit H+2. I did not really check all
  the instances but, from a random inspection, it seems the limit is now
  respected. 


Note: 
-----

Differences between the results of the OR Spectrum paper and the ones
obtained here can be due to a number of reasons, e.g.:
i. in the OR Spectrum version, some instances were not actually solved and,
therefore, the average per class was computed over less than 40 instances. This
was due to memory errors.
ii. the H+2 limit. Enforcing this limit changes the behavior of the corridor
method and, therefore, forces a different exploration mechanism.
iii. the hardware. Since the method is heuristic in nature, using a faster
hardware allows to perform a deeper exploration of the solution space (it would
be equivalent to increasing the time limit on the old hardware.) While this is
irrelevant for small instances (the CM finds the optimal solution for all the
instances of classes 3 and 4), it might be an issue for larger instances.
You can’t perform that action at this time.