The Metapopulation Genetic Algorithm

1
The Metapopulation Genetic Algorithm

• Lemon and Milinkovitch (2002). PNAS 99
  10516-10521
• METAPIGA (software) available at
  http//dbm.ulb.ac.be/ueg

2
Genetic algorithms

• Matsuda (1996), Lewis (1998), Katoh et al (2001),
  Foster (2001)
• A type of evolutionary computational method
• Mimic processes of biological evolution such as
  mutation, recombination, selection, and
  reproduction

3
Genetic algorithms

• Initial step generate a population of
  individuals (specific solutions to a problem)
• Individuals within a population are subjected to
  mutation and/or recombination
• Individuals allowed to reproduce with a prob.
  that is function of their relative fitness
  value
• Mean score of population improves over time

4
MetaGA essence of the procedure

• Coexistence of two or more populations
  interacting in a metapopulation setting
• Parallel searches allow significant
  inter-population variation in spite of strong
  selection
• But populations are not fully independent, they
  are forced to cooperate

5
MetaGA essence of the procedure
6
Consensus Prunning

• Communication between populations defines
  topological operators
• Assumption comparisons identify tree partitions
  that are correct (and should not be modified) and
  regions that still need to be resolved

7
Consensus Prunning
8
Consensus Prunning

• Random, ring, alternate ring, strict group
  consensus, majority group consensus, probability
  group consensus

9
Evaluation

• At the beginning of each generation, each tree is
  evaluated
• Evaluation does NOT involve optimization of
  branch lengths or other model parameters
• The logL is scored and recorded

10
Selection

• Rank by lnL Prob of leaving an offspring
  2(n-i1)/(n(ni)
• Tournament two indiv drawn at random, one
  offspring is produced from the best (with
  replacement)
• Replacement similar, but two copies of the best
  are produced (without replacement), sn times
• Improve only better trees than previous best
  will reproduce

11
Mutation

• Each individual, with the exception of the
  current best from each population is subjected to
  a single mutation
• Topological mutations SPR, NNI, taxa swap (TS),
  subtree swap (STS)
• Branch Length mutations (BLM) only on internal
  nodes
• Recombination (RCM)

12
Starting Trees

• Jacknifed NJ (JNJ) overlapping and
  non-overlapping
• Noisy NJ (NNJ)
• Random (not recommended)

13
Stopping rule

• The overall number of consensus partitions
  increases with search time
• Search stops when
• Best trees of all populations have the same
  topology, or..
• When all topological changes allowed have been
  attempted but no improvement was achieved
• User hits the stop button (or predefined of
  generations has been reached)