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 4. Solving the DRP 18<br />

4.3.1 A construction heuristic<br />

<strong>The</strong> initial solution is the basis for optimisation and as such, is an important part of any<br />

heuristic modelled to solve this problem. One goal of a construction heuristic is often<br />

to be fast as to allow time to be focused on other parts of the process. In the context of<br />

this DRP, several construction heuristics can be identified:<br />

Zero solution is a straightforward and easy solution to construct and consists of noth-<br />

ing but an empty schedule. This is a very fast (O(1)) solution to generate but is<br />

infeasible (over all shifts) due to constraint H4 on page 10.<br />

Random solution consists of randomly assigning doctors to shifts that are unassigned,<br />

until all shifts are assigned a doctor. This can be done by assigning each shift a<br />

random doctor, or, for each doctor, assigning that doctor to approximately n<br />

m<br />

random shifts. <strong>The</strong> random solution does not guarantee feasibility and is worse,<br />

in running time (O(n)), than the zero solution.<br />

Greedy solution is based on the concept of choosing what looks best, given a candi-<br />

date set and a selection function, here and now. That is, it bases its choice upon<br />

knowledge of the present and past shifts, but not the future ones as they have not<br />

yet been assigned a doctor. It makes no reconsideration of past choices, contrary<br />

to e.g. dynamic programming algorithms, that bases decisions on all previous de-<br />

cisions and as such, may reconsider the past. Several notions of the candidate set<br />

are used in the literature, but in the context of this project, I define the Local<br />

Candidate List as a list unique to each shift consisting of all doctors, ordered by<br />

their heuristic value when assigning them to the given shift. <strong>The</strong> greedy algorithm<br />

picks the best doctor from the candidate list and assigns it to the shift at hand.<br />

<strong>The</strong> basic version of the greedy is deterministic, i.e. given the same input, greedy<br />

always returns the same output. In an adaptive greedy algorithm (Resende and<br />

Ribeiro [20]), the doctor is randomly chosen from the local candidate list. To rule<br />

out the possibility of choosing the worst doctor for the shift, I introduce the Re-<br />

stricted Candidate List as the n best, measured by either a heuristic value or by a<br />

percentage, elements in the candidate list. Figure 4.2 on the next page illustrates<br />

various kinds of RCL parameters for a minimisation problem. Another option for<br />

the RCL list, is to define it as a list of ordered pairs, consisting of all unassigned<br />

shifts coupled with the doctor that has the highest heuristic value when assigned to<br />

the specific shift. This option was discarded due to the time required to construct<br />

the list being O(n · m) and the fact that the list would have to be reconstructed<br />

for each shift. <strong>The</strong> greedy solution does not guarantee feasible solutions, though<br />

the odds has improved vastly, from that of the random solution.

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

Saved successfully!

Ooh no, something went wrong!