Contact
office: +32 2 629 37 11
Emily Wang
Artificial Intelligence Laboratory
Vrije Universiteit Brussel
Pleinlaan 2
1050 Brussels
Belgium
curriculum vitae
letter
a4
|
|
Research
The Reciprocal Naming Game is an extension of the Naming Game
(Steels, 1995) that incorporates a signaling game (Crawford & Sobel,
1982), allowing the agents to consider personally selfish goals in
their choice and interpretation of utterances.
This model is aptly described as a shell game, where the context is
a set of shells, and a dealer has hidden a pea under one of
the shells. The hearer wins the game by choosing the shell that
contains the pea, which is information accessible to the speaker. The
speaker tells one name to the hearer, which may be a deception. If
the speaker confides the whereabouts of the pea truthfully, then it
has cooperated, acting as an informant to the hearer, and in the event
that the hearer wins, the reward is shared with the speaker. But if
the speaker defects and tries to deceive the hearer, acting as a shill
in support of the dealer, it receives a large reward if the hearer
chooses incorrectly. Action choices and the payoff function resemble
that of the repeated prisoner's dilemma, but with a linguistic twist
since messages are not guaranteed to be understood. The paper
presented at EVOLANG7 (see below) describes the game in more detail
and presents some different policies that may be used by the agents
for directing their cooperation. The social aspects of these policies
determine whether naming conventions can converge to stable
agreement.
For those familiar with modal logic, the entire Reciprocal Naming
Game can be represented succinctly as a system of eight possible
worlds, with indistinguishability relations labeled separately for the
speaker (S) and hearer (R).
Abstract (Wang & Steels, 2008):
We examine the social prerequisites for symbolic communication by
studying a language game embedded within a signaling game, in which
cooperation is possible but unenforced, and agents have incentive to
deceive. Despite this incentive, and even with persistent cheating,
naming conventions can still arise from strictly local interactions,
as long as agents employ sufficient mechanisms to detect
deceit. However, unfairly antagonistic strategies can undermine
lexical convergence. Simulated agents are shown to evolve trust
relations simultaneously with symbolic communication, suggesting that
human language need not be predicated upon existing social
relationships, although the cognitive capacity for social interaction
seems essential. Thus, language can develop given a balance between
restrained deception and revocable trust. Unconditional cooperation
and outright altruism are not necessary.
Publications & Presentations
-
Emily Wang
Lexical agreement between self-interested agents:
Language evolution in the context of
reciprocal altruism and the
cheater detection hypothesis
Master na master thesis, Vrije Universiteit Brussel.
pdf
figures
Presentation: Juried defense
Brussels (Belgium), 11 September 2007
pdf
-
Emily Wang
A model for learning the meaning and usage of numbers.
Senior thesis, Yale University.
pdf
cs.yale
The Past
During the 2006-07 academic year, I visited the VUB AI Lab on a
Fulbright fellowship sponsored by the U.S. Department of State.
As an undergraduate I worked in the Yale Social Robotics Laboratory.
For my senior thesis I implemented an algorithm for acquiring numeracy
from a combination of social imitation with both visual and auditory
input. This supervised learning algorithm was based on evidence from
cognitive science for the development of learned counting skills out
of an innate ability to subitize.
