Parameter Testing. The RPBSA algorithm requires the initialisation of 15 different input parameters. Some of these parameters were already explained in detail and values for those have been set after extensive test runs. The repair multiple τ has been the subject of Section 4.1.5.3 and is set to 0.75. Additionally, the initial values for parameters ω1 and ω2 of the dynamic operator selection (Section 4.1.5.2) have also been fixed to, respectively, (√n/2 + 0.005)% and 0.005%. 14xxxx://xxx.xxxxxxxxxxxxxxx.xxxxx.xx/RobinX/ An optimal combination of the 12 remaining parameters could potentially be obtained by performing a full factorial experiment. However, if every parameter would have only two possible levels, which is not the case, 4096 (212) additional experiments would have to be performed to cover each combination once and determine all main and interaction effects. In practice, since the parameters have more than two potential levels and a single replication is strongly discouraged due to the stochastic nature of the algorithm, parameter tuning is not done using a full factorial experiment design. Instead, parameter tuning is done one parameter at a time by performing test runs based on the default values presented by Xxxxxxxxxxxxxxx et al. (2006) and Xxx Xxxxxxxxxx and Vergados (2007). Consequently, interaction effects are not considered in this dissertation. More in depth parameter testing could be an interesting subject for future research. The results of the parameter tuning are summarised in Table 32. Parameters G, W , R, P and C seem to have a considerable impact on the runtime of the algorithm, but a minor influence on its qualitative outcome. Therefore, these parameters are set in a way that they return an algorithm with a manageable runtime. Appendix C presents the outcome of 3 5-minute test runs of every NLx instance for several parameters. Weight factor θ, cooling factors β and ζ and population parameters X and Y have been set to several values during this test but only marginal gains could be witnessed and are therefore set to their default value as determined by Xxx Xxxxxxxxxx and Vergados (2007). The same reasoning is followed for weight factor w, where the default value of 4000 turns out to be the best performing out of the 5 tested values. For the last parameter, T , the differences in outcome noticed between several values were significantly larger, as shown in Table 33. For instances of up to 8 teams, a temperature set to 400 generates better schedules...
