WP3: Language Evolution
Paul Vogt
Federico Divina
Tilburg University
Objectives (from Annex I)


… to design a population such that it is
capable of evolving one (or possibly more)
languages that enables them to optimize
cooperation.
A secondary objective is to design the
experiment such that the agents will
discover communication as a useful strategy
and find ways to use this strategy
effectively.
Tasks




Task 3.1 Define (…) the required set-up for evolving
language, learning how to use communication and
how to react properly on linguistic communication
(…). Year 1: M3.1
Task 3.2 Implement the code for under 3.1 defined
specifications and integrating the results achieved in
tasks 2.2 and 2.3. Year 2: D3.1
Task 3.3 Perform experiments with the system as
implemented in task 3.2. Started Year 2
Task 3.4 Report on the experiments performed.
Started Year 2
Overview






State of WP3
Language games
Preliminary results
Social learning of skills
Outlook final year
Conclusions
Language games
Form
“Cabbage”
Category
Category
Referent
Aspects of language learning

Establishing joint attention


Cross-situational learning


statistical co-occurrences across situations
Feedback


pointing
not reliable
Principle of contrast

associations with existing meanings lower initial score
Experiments


Aim: To test effect of learning mechanisms on
language development
Conditions:








Fixed controller (no individual learning)
Reproduction, but no evolution
Socialness gene randomly set
Possible actions: move, turn, pick-up, eat, mate, talk &
shout
Possible topics: features of one object
Fixed categories
Initial population size = 100
Simulated for 36,500 time steps (~100 NTYears)
Some statistics




Per time step: ~27 language games initiated
(total simulation ~1 million games)
~42% of games accompanied by pointing
gesture
~12% of games accompanied by feedback
signal
~50% of games no pointing, nor feedback
Divina & Vogt, Proc. EELC, 2006
Varying No. of Features
Vogt & Divina, Interaction Studies, in press
Excluding learning mechanisms
0,7
0,6
Accuracy
0,5
0,4
0,3
0,2
0,1
0
Standard
No Feedback No Principle
of Contrast
No Crosssituational
learning
No Pointing
Social learning
Assuming communication has evolved, how can
language be used to acquire new skills?
Example
A1
A2
T
E
E
L
T
h
f
M
T
R
“hungry,have-food, eat”
E
h
M
E
{h,f,E}
L
Example
A1
A2
T
E
E
L
T
h
f
M
T
R
“hungry,no-food,talk”
E
h
M
E
L
{h,f,E} {h,¬f,T}
Example
A2
T
E
h
M
E
L
{h,f,E} {h,¬f,T}
Example
A2
T
E
h
M
f
T
E
L
{h,f,E} {h,¬f,T}
Will it work?


Good question, we don’t know...
RL has (at least) 2 ways of deciding which
nodes to insert



Random insertion
‘Intelligent’ insertion
Our feeling is that second option could be
more effective and integrates language
evolution & social learning elegantly
Outlook final year


Integrating social learning (mostly done) – also
using ‘telepathy’
Performing experiments to



Focus of interest:





Improve model regarding accuracy
Evolve language that aids survival & social learning
Language diffusion
Emergence of dialects
Social learning
(Grammar)
Define language specific challenges
Conclusions




Made great progress
Language games work well beyond chance,
but could be improved
Social learning of skills defined,
implemented, but not integrated
Still much to do...
Descargar

WP3: Language Evolution