Reviewing the conclusions of Sections 126.96.36.199, 188.8.131.52 and 184.108.40.206 we may identify which stability parameters are useful and safe. For sparse datasets we concluded that rrlambda, cgwindow and cgdecay were helpful. For dense datasets it wasn't clear that any stability parameters were needed, and we noticed that cgwindow and cgdecay may need some tuning to prevent premature CG termination. While wmargin is successful in reducing occurrences of NaN values for non-rrlambda experiments, it becomes irrelevant if rrlambda is used. Our binitmean experiments suggest that binitmean should not be used with our IRLS implementation. Finally we noted that cgdecay is not necessary when cgwindow is enabled.
For the remaining IRLS experiments in this thesis, we will activate only rrlambda and cgwindow from the stability parameters. The values of these parameters will be chosen empirically in Section 220.127.116.11. The parameters have been added to our LR description tree in Figure 5.4.