Phat Phenotypes

1
Phat Phenotypes

• George Kampis
• Laszlo Gulyas
• Eotvos Uinversity Collegium Budapest

Research supported EU FP6 IST-5-033883-STP
2
What we do

• Phenotype based evolutionary modeling
• In an agent-based approach
• FATINT system, EvoTech project
• http//hps.elte.hu/kampis/EvoTech/

(niche construction)
3
(No Transcript)
4
(No Transcript)
5
(No Transcript)
6
Artificial slugs

• Genderless sexual reproduction
• Sexual selection based on similarity
• Placed in an evolutionary engine
• non-spatial, partial artificial ecology
• a single resource, energy
• full life-cycle (reproduction, aging, death)
• standard evolutionary operators mutation,
  crossing-over
• Plus....

7
Artificial slugs

• Genderless sexual reproduction
• Sexual selection based on similarity
• Placed in an evolutionary engine
• non-spatial, partial artificial ecology
• a single resource, energy
• full life-cycle (reproduction, aging, death)
• standard evolutionary operators mutation,
  crossing-over
• Plus....

• Changing interaction from phenotype plasticity

8
(No Transcript)
9
(No Transcript)
10
What we achieve

• Sponteneous separation of species in n-space
  (2004) phenotype driven open evolution

11
Simmons, Robert E. and Lue Scheepers. "Winning by
a Neck Sexual Selection in the Evolution of
Giraffe." The American Naturalist Nov 96 771-86.
12
Kampis, G. Gulyás, L. 2004 Sustained Evolution
from Changing Interaction, in Alife IX, MIT
Press, Boston, pp. 328-333.
13
(No Transcript)
14
What they are not
15
What they are (somewhat) like
16
What they are (somewhat) like
i.e. real body, real interaction
17
But the difference is...

• That (the phenotype of the) slug in FATINT
  feeds back to the evolutionary process itself
• i.e. interactor in the full sense
• Including an ability for flexible operations
  critical for niche construction, i.e. changing
  the environment for the same or other species

18
The problem of interactor modeling

• Preserving essential aspects of a body

• What is essential?
• I. Requirements for phenotype representation and
  the genotype/phenotype mapping
• II. Minimizing artifacts

19
I. Phenotype representation
20
E
G genes GP gene products E environment P
Phenotype
G
GP
P
21
P
P
( Gi, Ej )
Gi genes Ej environmental factors
P fat phenotype P narrow phenotype
22
Modeling of the fat phenotype

• Fat phenotype totality
• Is its modeling hopeless?
• Structural modeling, not concrete image
• Can be represented by permitting relational
  features
• RePast porous agents
• environment modulated artificial ontogeny

23
Relational properties 1.
Binary etc relations define them
24
Relational properties 2.
25
What you want to do...

• Phenotype induction by environmental components
• in variable length feature vectors
• That is, feature vectors are
• Context sensitive (depend on other f.v.-s)
• Bound to variation (external and internal)

26
Emergent Phenotypes
in natural and in model populations
Form Cause Type Point mutation endog. local P
henocopies exog. part global Epigenetic
change both part global Horizontal
adapt. both global Behavior change social global
27
Standpoint

• The study of interactors boils down to the study
  of
• fat phenotypes
• relational properties
• variable feature vectors

28
II. Minimizing artifacts
29
In a nutshell

• The giraffe has no feature vector

30
Making sense

• Model genotype/phenotype vectors (natural
  representation -?)
• Their representation in operators is arbitrary

31
Arbitrariness

• Real world example
• Similarity based sexual selection
• Genotype similarity phenotype similarity
• Similarity markers, e.g. odor, shape etc.
• Modeling similarity with feature vectors
• Metric distance
• Pattern matching
• Focal dimensions

32
An analysis...
33
Metric model

• The threshold T defines an L-dimensional
  hyper-sphere of radius T around the phenotype of
  agent i containing the possible phenotypes of
  potential partners. On the other hand, the
  hyper-cube Vmin, VmaxL holds all theoretically
  possible phenotypes in L dimensions. The
  likelihood of finding partners with ?(i, j)ltT is
  expressed as the ratio of the volumes of the
  above hyper-sphere and hyper-cube. For the sake
  of simplicity, lets consider the volume of a
  hyper-sphere of unit diameter over the volume of
  a hyper-cube of unit side length, which is an
  obvious upper limit. This ratio is given by the
  following formula (Singmaster, 1964)

34
Summary, metric model

• The metric model tends to split species as L
  increases
• The giraffe has no n-dimensional distance
  function.

35
Pattern matching I.

• where ?() is the membership function and ? is
  the one-dimensional, pair-wise matching operator.
  Naturally, it is often convenient to normalize
  similarity values, especially if L is not fixed
  during the execution of the model.

But, lets consider two phenotype vectors A and
B, for which the required action is feasible.
Then introduce an additional dimension in which
the two individuals do not match. This
necessarily re-scales the similarity score of the
vectors, which may render the action unfeasible
again. If the action in question was similarity
based mating, then this means that the
introduction of the novel dimension may
artificially split existing species.
36
Pattern matching II.
An interesting consequence of this puzzle
matching is the spontaneous appearance of
sexes. The requirement that surviving members
of a population should be able to perform an
action whose feasibility is defined as above will
eventually lead to the formation of complementary
group pairs where members of one group can
perform the action with members of the
corresponding group only. Unfortunately,
however, this leads to a fatal artifact it
creates a pronounced drive towards taking middle
values (i.e., (VminVmax)/2) at all k?1, L
locations. The group of individuals with middle
values can always perform the given action
(i.e., the group is self-complementing).
37
Summary, pattern matching

• Pattern matching as in I or II (and other
  methods) tends to split or merge species
• The giraffe has no n-dimensional pattern
  matching operator.

38
Focal dimensions

• For the sake of simplicity, let the first T
  positions specify the focal dimensions. As
  Vmax-Vmin can be greater than L, we assume that
  index values are understood as modulo L.

It is important to realize, however, that this
similarity operator is necessarily non-symmetric.
This could result in a non-intended agglomeration
of several species. The non-focal dimensions face
no specific selection pressure and thus the
groups of individuals that diverge widely in a
common set of focal dimensions will be
nevertheless likely to share values here
39
Summary, focal dimensions

• Focal dimensions tend to merge species.
• The giraffe has no focal dimensions in n-space.

40
But that is not quite true

• There should be a way to augment n-space with
  operators that mirror the ecological solution
  species experience similar hitchhikers yet
  maintain different features
• ...Or see current research on tuple space

41
Conclusions, interactor modeling

• I. Phenotype representations require relational
  structures (porous agents)
• II. Artifact minimization possible?

42
References

• Kampis, G. 2002 A Causal Model of Evolution,
  Proceedigs of. SEAL 02 (4th Asia-Pacific
  Conference on Simulated Evolution And Learning),
  Singapore, pp. 836-840.
• Kampis, G. Gulyás, L. 2004 Emergence out of
  Interaction A Phenotype Based Model of Species
  Evolution, Proceedings of IWES 04, pp. 83-89.
• Kampis, G. Gulyás, L. 2005 Sustained Evolution
  from Changing Interaction, in Alife IX, MIT
  Press, Boston, pp. 328-333.
• Kampis, G. Gulyas, L. 2006 Full Body The
  Importance of the Phenotype in Evolution, ECO
  Workshop, ALifeX Conference
• Kampis, G. Gulyas, L. 2006 Emergence as a
  Relational Property in Societies of Agents , AAAI
  Fall Symposium, Interaction and Emergent
  Phenomena in Societies of Agents, Arlington, VA.
• Project Web Site
• http//hps.elte.hu/kampis/EvoTech/ET.htm

43
(No Transcript)