Communication is a Coordination Problem

1
Communication is a Coordination Problem
Les Gasser Graduate School of Library and
Information Science University of Illinois at
Urbana-Champaign gasser_at_uiuc.edu
IBERAMIA 2004 Puebla, Mexico
Joint work with the UIUC Language Evolution and
Computation GroupBrant Chee, Marc Esteva, Kiran
Lakkaraju, Samarth Swarup, Jun Wang See also
http//www.isrl.uiuc.edu/amag/langev
2
This is a talk about adaptive, coordinated
communication, and specifically about autonomous
development of languages in artificial agent
societies.
3
Why study adaptive language?

• Intrinsic scientific interest
• Human language emergence and evolution
• Resource Description/Discovery, web services,
  etc.
• Vocabulary coherence in IR P2P IR
• HCI- learning to talk with robots, where people
• adapt to robots too
• 7. Bioinformatics representation in biological
  systems biosystems as languages

4
Conclusions

• Inaccessible information drives communication
• Communication requires public language.
• Public language is a decentralized coordination
  problem
• Public language must adapt
• Hypothesis computational agents can solve the
  adaptive collective public language problem
  autonomously.

5
Why communicate?
6
Model Problem 1Poison Mushroom Awareness(A.
Cangelosi)
7
Model Problem 1Poison Mushroom Awareness
8
Option 1 Better poison detection
9
Option 2Communication
Communicate with anagent that can tell
..communication enables agents to exploit a
division of labor
10
Why Communicate?
Information/Capability is Inaccessible ?? Which
mushroooms are poison ??
? Cost ? Motivation
Get it fromanother agentCommunicate
Create it myself Poison detectors
11
Model Problem 2 Coordinated Search (inspired by
Werner Dyer, 1992)
Actor Apply Operators Move All brawn, no
brains
RandomSearch
Goal
12
Model Problem 2 Coordinated Search
Actor Apply Operators Move All brawn, no
brains
RandomSearch
Observer Sense/Observe Evaluate Compute Best
Operators All brains, no brawn
Goal
13
Division of Labor Performance Improvement Goal
(Directed gt Random)Create Interdependency
14
Why communicate?
Information/Capability is Inaccessible ?? Moving
to goal ?? ?? Knowing best move ??
? Cost ? Motivation
Get it fromanother agentCommunicate
Create it myself Learn to MoveLearn to
sense/reason
15
Interdependency CreatesNeed for Communication
16
(Recall We have many agents) (populations)
17
Summary Coordinated Search

• Communication enables agents to exploit division
  of labor
• Coordinated Search is a general model of MAS
  situations
• Why communicate?
• Inaccessibility Observers information is
  inaccessible to Actor
• Motivation Both agents want performance
  improvement
• Lowest Cost Alternative building communication
  capability is cheapest alternative (else Actor
  would learn to sense reason Observer would
  learn to apply operators)
• These create an interdependency need for
  communication

18
Why Communicate?(General Model)
Information Inaccessibility Problem
Exists Critical information is differentially
accessible Communication is Highest-Utility
Solution Building communication capability is
alternative with highest amortized utility for
all participants (else other avenues would be
developed, e.g., improved sensing/reasoning)
19
Aims of communication

• Assuring a transfer of information across
  spacetime
• 2. Changing the mental state of the receiver in
  a predictable way across spacetime

Need a theory of how communication can modify the
knowledge state of agents Basic question what
kinds of communication acts guarantee predictable
knowledge states among agents (and hence enable
coordination)?
20
Example Common Knowledge
Coordinated Attack Two allied generals are
stationed on opposite hills. In order to
successfully attack the enemy in the valley, they
must attack together. General A sends a
messenger to General B, with the message Attack
at dawn!
21
Example Coordinated Attack
Coordinated Attack Two allied generals are
stationed on opposite hills. In order to
successfully attack the enemy in the valley, they
must attack together. General A sends a
messenger to General B, with the message Attack
at dawn!
A
B
knows(A, ltattack at dawngt)
knows(A,knows(B,ltattack at dawngt))?
MSG Attack at dawn!
22
Attaining Common Knowledge(Halpern and Moses,
1984, 1990)
When communication is not guaranteed it is
impossible to attain common knowledge Common
knowledge requires simultaneous (coordinated)
action, which cannot be guaranteed. Weaker,
useful variants exploit eventual and time-bounded
communication, and ?-common knowledge. What
about coordination in communication?
23
Understanding Coordination inCommunication
24
Understanding Coordination inCommunication
Actor
Random
Search
Apply Operators
Turn Left
Move

All brawn, no brains

Directed
Search
Observer
Sense/Observe
Evaluate
Goal
Compute Best Operators

All brains, no brawn

25
Shannonsmodel
Understanding Coordination inCommunication
Actor
Random
Search
Apply Operators
Turn Left
Move

All brawn, no brains

Directed
Search
Observer
Sense/Observe
Evaluate
Goal
Compute Best Operators

All brains, no brawn

26
Coordination in Communication
Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
27
Coordination Needed (a b)
Coding Coordination(over encoding/decoding space)
Signal Coordination(over transmission noise)


Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
28
Coordination Approaches (a)
Signal Coordination (a)(over transmission noise)
100110110101
100110110101

100110110101
100110110101
Solutions redundancy error correction noise
reduction All employ global transmitter-receiver
coordination
29
Coordination Approaches (b)
Coding Coordination (b)(over encoding/decoding
space)


Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
Solution inverse signal encoders-decoders
(globally-coordinated)
30
What about semantics (c)?
Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
Semantic Coordination(over meaning space) What
does Turn Left mean to each?
31
What about semantics (c)?
Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
Actors world
Observers world
pos(X,Y) orientation(Z?) ? pos(X,Y) orientation(Z?
?)
Coord1(a) Coord2(b) direction(dminutes)
Semantic Coordination (c)(over meaning
space) What does Turn Left mean to each?
32
Summary

• Levels of coordination in communication
• Signal Coordination (over transmission noise)
• Coding Coordination (over encoding/decoding
  space)
• Semantic Coordination (over meaning space)
• Global approaches assume we can a) anticipate
  conditions and b) control both ends
• From now on we focus on 3

33
What about semantics (c)?
Turn Left
Turn Left
100110110101
100110110101
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
Actors world
Observers world
pos(X,Y) orientation(Z?) ? pos(X,Y) orientation(Z?
?)
Coord1(a) Coord2(b) direction(dminutes)
Semantic Coordination(over meaning space) What
does Turn Left mean to each?
34
What about semantics (c)?(assume away encoding
transmission coordination)
Turn Left
Turn Left
Actor Apply Operators Move
Observer Sense/Observe Evaluate Compute Best
Operators
Actors world
Observers world
pos(X,Y) orientation(Z?) ? pos(X,Y) orientation(Z?
?)
Coord1(a) Coord2(b) direction(dminutes)
Semantic Coordination(over meaning space) What
does Turn Left mean to each?
35
Semantic CoordinationThrough Language
WO and WA may or may not be the same
36
Semantic CoordinationThrough Language
Private conceptualizations(inaccessible across
agents)
Assume they are the same
37
Semantic CoordinationThrough Language
Turn Left
Turn Left
Public Language (accessible to both)
Private conceptualizations(inaccessible across
agents)
38
Origin of the public language?
Present Human design for all participating
agents not adaptive excludes non-participating
agents limits autonomy.. Future Agents
themselves create language So, what is language
and how can agents create it?
39
Language Elements
Adaptive, Collective(no single-agent language)
Concepts (what to talk about) Symbols
(arbitrary signs) Grammars compositionality
(infinite use of finite means)
40
Language as Mapping
Language can be viewed as a mapping between
(possibly complex) meanings and (possibly
complex) expressive forms. Forms can be
publically mobilized meanings cannot
(inaccessibility of meanings)
41
Communication
42
Semantic CoordinationThrough Language
Turn Left
Turn Left
Public Language (accessible to both)
Q
Meaning to Form Mapping
Form to Meaning Mapping
P
Private conceptualizations(inaccessible across
agents)
43
Examples ( as probability matrices)
Forms
P
Meanings
Production
Forms
Q
Meanings
Interpretation
44
Examples (Imperfect communication)
Forms
Synonymy
P
Meanings
Forms
Synonymy
Ambiguity
Q
Meanings
Poverty
45
Example (Perfect Communication)
Forms
P
Meanings
Forms
Q
Meanings
46
Indexical and Compositional Languages
When are intensional, giving rules for
mapping (sets of) meanings to (sets of) forms
(i.e., a grammar) the language is compositional,
and may be infinite and/or recursive.
Compositional languages may improve efficiency
(cf. work of Nowak et al.) Infinite use of
finite means
47
Communicating Agents
Agenti
Agentj
Forms
Forms
Meanings
Meanings
P
Q
Production
Interpretation
Forms
Forms
Meanings
Meanings
Q
P
Interpreation
Production
48
Communicating Agents
Agenti
Agentj
Forms
Forms
4 Mapsto Coordinate
Meanings
Meanings
P
Q
Production
Interpretation
Forms
Forms
Meanings
Meanings
Q
P
Interpreation
Production
49
Population-Level Coordination Issues
Communicability Probability that sentence in Li
is understood by speaker of Lj Efficiency Efficie
ncy of production (Li), transmission,
interpretation (Lj) Expressiveness Ability of
Li, Lj to express critical concepts Across
a population of agents (? Li,Lj)
50
Language Coordination
L1Ln is the Language Cloud
Li - Lj lt ?
Limits? Mechanisms?
51
Many Interacting Systems
Collective adaptive language design (aka
language evolution) is hard because it involves
interactions among several interacting complex
systems..
52
3 interacting units/timescales of adaptation
Unit
Determines
Over period
AgentArchitecture(map structure update)
Space of languages learnable by agents (aka
Universal Grammar)
Evolutionary Time
LanguageCloud(languages inthe population)
Cultural (glossogenic) Time
Space of current learning samples
IndividualLinguisticKnowledge (state of maps)
Statistical distribution of communicability
(selection pressure)
AgentLifetime (Critical Period)
53
Typical approaches inspired by human language
evolution models
54
Simple experiment(Java / RePast 2.2)
Observer Perception(A) ? Move(A) Move(A) ?
Message(O,A) 3 possible moves 8 possible message
forms is a 3 x 8 probability matrix
55
Experimental Issues
56
Initial Conditions
57
Algorithm
58
Results
Synonymy
Accuracy
Ambiguity
59
Results
Tasks completed per 200 cycles Blue no
communication (random search) Red adaptive
communication (directed search)
60
Results
Average messages per task completion per 200
cycles Blue perfect communication Red
adaptive communication (directed search)
61
Take-Home Points

• Communication makes available critical,
  inaccessible information.
• Communication requires public language.
• Public language is a decentralized coordination
  problem balancing (at least) communicability,
  efficiency, and expressiveness.
• Public language must adapt to changes in
  information, criticality, or system economics
• Hypothesis computational agents can solve the
  adaptive collective public language problem
  autonomously.