New Search Technologies

1
Lecture 4

• New Search Technologies

2
A new search paradigm emerged

• Cannot afford to get hung up finding every tree
  on huge islands
• Do not allow tree buffers to fill up
• Hold only few trees at each length in each
  replicate
• Pool results and pass later to TBR

3
A more powerful approach Parsimony Ratchet

• Kevin Nixon (1999) - Parsimony Ratchet
• Method for more thorough exploration of tree
  space than RASTBR
• Branch-swapping in suboptimal areas may lead to
  more optimal solutions

4
A more powerful approach Parsimony Ratchet

• Generate starting tree
• Branch-swap until an optimum is reached
• Save trees
• Reweight some 5-25 of characters chosen at
  random
• (Can be either up or down weighted)
• Branch-swap until new optimum reached
• Reweight characters back to original weights
• Branch-swap again
• Continue for 50-200 cycles

5
Parsimony Ratchet

• Uses perturbations to explore tree space
• Move through suboptimal areas

6
Goloboff (1999)

• RAS TBR will probably never succeed with large
  data sets
• Parsimony ratchet is better
• But, entirely new approaches are needed for data
  sets of 150 or more taxa

7
Goloboff (1999)

• Best initial strategy is to conduct aggressive
  initial searches
• Drill down to shortest solutions quickly
• Dont worry about number of equally parsimonious
  solutions
• These can always be calculated later
• In fact, entirely unnecessary to find all equally
  parsimonious trees on every island
• Consensus will be the same if most islands in
  tree space are sampled even with 1 or 2 trees each

8
Goloboff (1999)

• Imagine tree with 500 taxa
• Taxa arrayed in 10 sectors of 50 each
• Each sector has its own local optimum
• Local optima more or less independent of one
  another
• Global optimum - all sectors must be optimal
  simultaneously!
• Search strategy must implicitly deal with
  composite optima

9
Tree Fusing (Goloboff 1999)

• Exchange sectors between multiple trees
  containing identical sets of taxa
• Each sector optimal on at least one tree
• Recombine sectors to find shorter global solutions

10
Sectorial Searches

• Break up large problem into sectors
• Analyze each separately
• 3 ways to determine sectors

11
1- Random Sectorial Searches

• Random Sectorial Searches (RSS)
• Analyze on particular sectors at random
• Can set size of sectors in advance
• Conduct branch-swapping on sector
• If shorter one found, insert back in tree
• Follow local, sectorial searches with global
  search (TBR) to insure global optimality

12
2- Consensus Sectorial Searches

• Focus on problem areas
• Areas of poor resolution (polytomies)
• Typically smaller than random sectors

13
3- Exclusive Sectorial Search (XSS)

• Divide tree into specific number of
  non-overlapping (exclusive) regions
• Entire tree is analyzed

14
Tree-Drifting

• Resembles Parsimony Ratchet
• Different way of moving around in tree space
• Accept suboptimal solutions, but not too
  suboptimal!
• Relative fit difference between trees A and B
• RFDAB (F - C)/F
• F sum characters that differ between A and B
  that fit tree A better
• C sum of characters that differ between A and B
  that fit tree B better
• Better than raw length difference because it
  accounts for actual conflicts between characters

15
Tree Drifting

• Reject suboptimal solution if
• RFD gt Z
• Z X / (F J - C)
• X is random number between 0 and 99
• J is length difference
• F and C are calculated for tree being swapped on
  and candidate tree
• Drift through tree space accepting trees with
  probability based on relative fit difference of
  characters on tree in memory and candidate tree
• Stochastic component to acceptance criterion

16
TNT ProgramPablo Goloboff, Kevin Nixon, Steve
Farris

• Three search modes
• Implicit Enumeration (exact searches)
• Traditional Search Methods
• random addition sequences
• TBR or SPR branch-swapping
• New Technology Searches
• Sectorial searches
• New implementation of parsimony ratchet
• Tree-drifting
• Tree fusion

17
TNT - New Technology Searching

• Can pass trees from one algorithm to another
• Alternate cycles of TBR branch-swapping,
  sectorial searches, parsimony ratchet and
  tree-drifting
• Typically finish with tree-fusing

18
TNT - Driver Program

• Overall control program for conducting searches
• Sort of a parsimony Expert System
• Can structure searches in various ways depending
  on stage of analysis
• Quickly bore down to find shorter trees
• Once most parsimonious scores are repeating,
  shift to examination of structure of consensus
  tree
• Other advanced features

19
References

• Goloboff, P.A. (1999). Analyzing large data sets
  in reasonable times solutions for composite
  optima. Cladistics 15 415-428.
• Nixon, K.C. (1999). The parsimony ratchet a
  new method for rapid parsimony analysis.
  Cladistics 407-414.