Demo THSim

Demo THSim
Tutorial modelling language evolution
Paul Vogt
Modelling language origins and evolution
IJCAI-05
THSim - Talking Heads simulation tool
http://www.ling.ed.ac.uk/~paulv/thsim.html
Modelling language origins and evolution
IJCAI-05
Discrimination
• World is a collection of objects (shapes on whiteboard)
– Represented as features: Red, Green, Blue, Shape (A),
X, Y
– Context = a set of objects on white board
– Topic = one particular object
• Robots want to build a set of meanings
• Meaning is a region represented by a prototype
– A particular colour, area and location
• The category of every object is the region represented
by its nearest prototype
• An object is discriminated if its category is different
from all the others in the context
• If discrimination fails, a new category is constructed
by taking the topic’s features to form a new prototype
Modelling language origins and evolution
IJCAI-05
Simplified example
CONTEXT:
A=(0.1, 0.3)
B=(0.3, 0.3)
C=(0.25, 0.15)
A
B
b
a
ROBOT’S PROTOTYPES:
a=(0.15, 0.25)
b=(0.35, 0.3)
C
A is categorised as a
B is categorised as b
C is categorised as b
A is discriminated
B and C are not
Modelling language origins and evolution
IJCAI-05
Guessing game
• Speaker produces an utterance to name the
topic.
• Hearer guesses the reference of the topic by
searching its lexicon for most likely
interpretation.
• Speaker provides corrective feedback on the
outcome.
• Agents adapt lexicon:
–
–
–
–
Speaker may produce new word.
Hearer may adopt utterance.
Successfully used associations reinforced.
Unsuccessful associations inhibited.
Modelling language origins and evolution
IJCAI-05
Demo I
• Showing the workings.
– Population size: 2
– Used features: Red, Green, Blue.
– World: Fixed set of 12 colours.
– Nr of games: 500.
Modelling language origins and evolution
IJCAI-05
Demo II
• Studying the effect of perceptual noise
(pNoise). Each agent sense the objects’
features with added noise:
• Each feature fi becomes fi’ = g(,x)  fi
– g(,x) = 2-G(,x)) if x<0
G(,x)
otherwise.
– x is a random value between -0.5 and +0.5.
– G(,x) is Gaussian with standard deviation  around
the mean 0
–  = pNoise
• Varying pNoise with same settings as in
demo I shows that system robust to noise, if
noise is not too strong.
Modelling language origins and evolution
IJCAI-05
Demo III
• Settings as in Demo I, but instead of fixed set
of colours, the colours are generated as
random Red, Green and Blue values.
(Unselecting fCol)
• Unstructured is more difficult to learn.
Modelling language origins and evolution
IJCAI-05
Demo IV
• Settings as in Demo I. Adding features Shape
(A), X and Y one by one.
• Shows that system is robust under increasing
complexity of meanings, though learning
takes longer.
Modelling language origins and evolution
IJCAI-05
Demo V
• Back to settings in Demo I. Increasing
population sizes and nr. of language games.
• Again language converges, though learning
takes longer.
Modelling language origins and evolution
IJCAI-05
Demo VI
• Settings as in Demo I.
• Varying game type, comparing
– Guessing game (based on corrective feedback)
– Observational game (based on joint attention)
– Selfish game (based on cross-situational statistical
learning – guessing, but no corrective feedback)
• Default update rule for association score sij between
meaning mi and word wj:
– sij = sij + (1-) Xij
• Where  is a learning rate and Xij=1 if used successful, and
Xij=0 if used unsuccessful.
• Selfish game works ‘only’ on Bayesian statistics, i.e.
– sij = P(mi|wj) = use(mi & wi)/use(wi)
• Where P(m|w) is the conditional probability that m is
observed when w occurs, and use(x) is the number of times
x is used.
Modelling language origins and evolution
IJCAI-05