Ratchet
From TNT
The ratchet in TNT is slightly different from the ratchet as originally described by Nixon (1999), and as implemented in Nona/Pee-Wee. The ratchet consists of two phases, perturbation and search, which are sequentially repeated. In either phase, the ratchet in TNT does not look for multiple trees (in Pee-Wee/Nona it was necessary to set the number of trees to save in each phase, but the best number is always just one); multiple trees can be found only as the search phase finishes. The ratchet always uses TBR branch swapping, and it alternates the search phase (S) with three types of perturbation phase: original weights (O), upweighting (U), and deleting (D). Thus, the cycles will be: O,S,U,S,D,S,O,S,U,S,D,S,O… etc. Any of the O, U, or D phases can be skipped (from the Ratchet settings dialog, or with the ratchet command). During the perturbation phase, rearrangements of score equal or better than the tree being swapped are always accepted, untill a certain number of rearrangements have been made (or untill a certain percentage of the total swapping on the tree has been completed). This seems to provide better results than the original ratchet. The "autoconstrained" option calculates the (strict) consensus of the previous tree and the tree resulting from the rearrangement phase, and during the subsequent search phase, only rearrangements that do not violate monophyly of the shared groups are done. The rationale for this is that improvements are likely to be found in areas that have changed during the perturbation phase. Every certain numer of constrained cycles, an unconstrained search phase is done (to insure global optimality). If this is used in combination with user-defined constraints, both are combined.
