27.07.2013 Views

The Doctor Rostering Problem - Asser Fahrenholz

The Doctor Rostering Problem - Asser Fahrenholz

The Doctor Rostering Problem - Asser Fahrenholz

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Chapter 7. Tests, results and discussion 46<br />

7.2.2 Metaheuristic parameter effects<br />

In this section I will discuss the effects of changing the metaheuristic parameters. In the<br />

tables presented in this section, the average results of SA is the results of giving SA the<br />

output from GRASP. As such, there are a much higher number of SA runs than there<br />

is GRASP.<br />

<strong>The</strong> RCL parameter<br />

<strong>The</strong> RCL parameter is fundamental to GRASP and allows the metaheuristic to ran-<br />

domise the construction of the initial solution in each iteration.<br />

Imps Z(S) V(S)<br />

GRASP 1 17,1875 3,5625 25,9375<br />

GRASP 100 14,625 392,75 19,3125<br />

GRASP 50 13,5 5,625 21,125<br />

Table 7.4: RCL effect: Test 0<br />

Imps Z(S) V(S)<br />

GRASP 1 19,625 5 32<br />

GRASP 100 15,3125 363,6875 27,875<br />

GRASP 50 13,625 6,125 27,5<br />

Table 7.5: RCL effect: Test 1<br />

Table 7.4 shows how, against suspecion, a more random RCL parameter, GRASP 100<br />

(393,19) actually results in a lower V(S). It does come with a cost, namely the high<br />

Z(S). In this case, I suspect that a more balanced result, such as for the GRASP 50<br />

(6,21), would be more preferable to the medical practice. <strong>The</strong> same goes for test 1,<br />

where GRASP 100 (364,28) produces a lower V(S) than GRASP 1 (5,32). This is, also<br />

described earlier, a key observation, which is not that far from what Resende and Feo<br />

[19] says. A fully greedy RCL may produce good mean values, but when we increase<br />

the randomisation, we may produce better results.<br />

<strong>The</strong> conclusion that a fully greedy RCL parameter is not optimal is supported by test<br />

8 and 9, shown in table 7.7 and 7.6 respectively. In these tables we see an obvious<br />

reduction in V(S) when going from fully greedy to fully random RCL parameter, even<br />

on test 9 where there are no night shifts, evening shifts or wishes.<br />

Imps Z(S) V(S)<br />

GRASP 1 7,6875 3,375 8,125<br />

GRASP 100 8,75 7,5 1<br />

GRASP 50 6,0625 2,875 3,375<br />

Table 7.6: RCL effect: Test 9<br />

Imps Z(S) V(S)<br />

GRASP 1 15,5 4,9375 19<br />

GRASP 100 11,0625 185,0625 9,625<br />

GRASP 50 14,625 4,625 14,9375<br />

Table 7.7: RCL effect: Test 8<br />

Over all tests, the average values of the solutions grouped by RCL parameters are shown<br />

in table 7.8 on the following page. In this table we can see that the average value of<br />

Z(S) increases significantly when the metaheuristic manage to decrease V(S).

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!