Strategic Considerations in Agent Dialogue games
Christos HadjinikolisSupervisors: Dr. Sanjay Modgil, Dr. Elizabeth Black, Prof. Peter McBurney
2
Reaching Agreements
• Negotiation dialogues– A bargain over the division of some resource– Negotiation is intended to aim at compromise
• Deliberation dialogues– Decide the action or the course of actions that
they should adopt in order to bring about some task
16/11/2011 King's College London Department of Informatics
3
Persuasion Dialogues
• How do persuasion dialogues fit into negotiation and deliberation?– “A participant tries to convince the other to accept
a proposition that the last does not currently endorse”
– Persuasion dialogues are essentially the means through which we resolve conflicts of opinion
16/11/2011 King's College LondonDepartment of Informatics
4
Persuasion Dialogues
• How does this form of dialogues fit into the picture?
16/11/2011 King's College London Department of Informatics
• How can such conflicts appear in these types of dialogues?
5
Example
16/11/2011 King's College London Department of Informatics
A
B
6
Persuasion Dialogues
• Deliberation dialogues:– In order to agree on accepting the proposed course
of actions, the proposing party needs to first convince its interlocutor on the acceptability of the claims about beliefs on which the proposition relies
• Negotiation dialogues:– “In a negotiation dialogue it is the reject move that
shows that there is a conflict between the preferences of an agent and the offer that it receives” H. Prakken, “A Protocol for Arguing about Rejections in Negotiation”
16/11/2011 King's College London Department of Informatics
7
Persuasion Dialogues
• Employed as embedded dialogues or sub-dialogues– Resolve conflicts– Optimise the duration of a dialogue and allow for
rationalising about a response
16/11/2011 King's College London Department of Informatics
8
Strategising
• Strategising in dialogues:– The participants have self interested objectives– Dialogues do not have objectives!• As McBurney & Parsons explain in their work on “Games
That Agents Play”: “it makes no sense to talk about the goals of a dialogue since the ones who actually have goals are the participants”
16/11/2011 King's College London Department of Informatics
9
Strategising
• What are the prerequisites of Strategising:– Information about one’s opponent• Abilities• Objectives• Its knowledge
– Opponent modeling
16/11/2011 King's College London Department of Informatics
10
Opponent Modelling
• How can such a model be built?– Through collecting information during the course of a
dialogue game or through a series of dialogue games– Provided by a an external source– Through observing other dialogue games as a third
party agent– Goals: Through observing its actions in the
environment, or even during the course of dialogue games in general, either as a participant or as an observer
16/11/2011 King's College London Department of Informatics
11
Opponent Modelling
• How can such a model be represented?– Our work relies on the employment of
argumentative systems for dialogue– An opponent model can be expressed in the same
that an agent’s own beliefs are, through an argumentation framework
16/11/2011 King's College London Department of Informatics
12
Opponent Modelling
16/11/2011 King's College London Department of Informatics
An agent’s own KB Its opponent’s KB A
B
C
D
A
B
C
D
13
Opponent Modelling
• We rely on the employment of an argumentative system for dialogue, but based on modelling actual knowledge!
• Why?– Because we believe that otherwise it is difficult to account
for the dynamic nature of dialogues which can only be captured though the underlying logic
– This concerns the possibility of new arguments being instantiated in the course of a game
16/11/2011 King's College London Department of Informatics
14
Opponent Modelling
• For this reason we rely on ASPIC+– Why?• It explicitly models the logical content and structure of
arguments, while at the same time it accommodates many existing logics for argumentation.
16/11/2011 King's College London Department of Informatics
15
The ASPIC Framework (2006)
• Developed by:– Leila Amgoud– Martin Caminada – Claudette Cayrol– Marie-Christine Lagasquie-Schieux– Henry Prakken– Gerard Vreeswijk
16/11/2011 King's College London Department of Informatics
16
ASPIC
• Relied on Dung’s framework and added to its expressiveness:• Described a general logical language L• Differentiated between strict and defeasible rules• Defined arguments with respect to their logical structure
– Logical premises– Rules– Conclusion
• Differentiated between the conflicts between arguments– Undercutting attacks– Rebutting attacks
16/11/2011 King's College London Department of Informatics
17
From ASPIC to ASPIC+
• Added another form of attack : undermining• From the notion of contradiction between formulas φ and ¬ φ, to an abstract relation of contrariness between formulas
• Distinguished between 4 types of premises– axioms, ordinary, assumptions, issues
• Attacks succeed as defeat relations based on:– preference orderings on arguments which in turn are based on:
» Priority orderings over defeasible rules and premises
• Unlike ASPIC, Prakken’s ASPIC+ showed satisfaction of Caminada’s and Amgoud’s rationality postulates when accounting preferences
16/11/2011 King's College London Department of Informatics
18
Our idea
• Knowledge representation:– agent’s ’s Knowledge base– : <, , , ... , >
» n: number of agents in the environment
• All agents share the same logical language L and contrary relation definition
16/11/2011 King's College London Department of Informatics
19
Our idea
• , ,, >
What believes is agent ’s :– Premises (– Pre-ordering over premises (– ()– Pre-ordering over t)– Goals ()
16/11/2011 King's College London Department of Informatics
20
Multi-Agent Knowledge Base
16/11/2011 King's College London Department of Informatics
21
The proposed approach
• The information gathered about the interlocutor• Based on a set of protocol rules
– Backtracking– Commitment stores
A strategy function is employed in order to choose from a set of legal arguments, the most suiting one with respect to one’s objectives
16/11/2011 King's College London Department of Informatics
22
The strategy function
16/11/2011 King's College London Department of Informatics
23
Example
• Let’s assume a dialogue protocol for grounded semantics:– Backtracking is allowed – Commitment store– Knowledge about what the interlocutor believes is 100% correct– Under the grounded protocol rules the proponent is not allowed to
repeat the same move twice while opponent can.
16/11/2011 King's College London Department of Informatics
24
Example: A dialogue game for grounded semantics
16/11/2011 King's College London Department of Informatics
A
B
C
D
E
F
G
H
25
Example: A dialogue game for grounded semantics
16/11/2011 King's College London Department of Informatics
A
B
C
D
E
F
G
H
04/19/2023 King's College London, Department of Informatics
26
A
B
C
D
E
F
G
H
Example: A dialogue game for grounded semantics
04/19/2023 King's College London, Department of Informatics
27
Agi’s Knowledge Base
s=> a p => ¬sr => a p => ¬ap => q q => ¬r ��s p r
X
Y
Z K
Example: A dialogue game for grounded semantics
S𝒊 𝒊 S𝒊 𝒋
T
W
F
04/19/2023 King's College London, Department of Informatics
28
Agi’s Knowledge Base
s=> a p => ¬sr => a p => ¬ap => q q => ¬r ��s p r
X
Y
Z K
Example: A dialogue game for grounded semantics
S𝒊 𝒊 S𝒊 𝒋
T
W
F
04/19/2023 King's College London, Department of Informatics
29
Agi’s Knowledge Base
s=> a p => ¬sr => a p => ¬ap => q q => ¬r ��s p r
X
Y
Z K
Example: A dialogue game for grounded semantics
S𝒊 𝒊 S𝒊 𝒋
T
W
F
win!
¬a ,¬a => g
…
04/19/2023 King's College London, Department of Informatics
30
Agi’s Knowledge Base
s=> a p => ¬sr => a p => ¬ap => q q => ¬r ��s p r
X
Y
Z K
Example: A dialogue game for grounded semantics
S𝒊 𝒊 S𝒊 𝒋
T
W
F
¬a ,¬a => g
Y
K
Y
Z K
r => ar
��q => ¬rr => a, p => qr, p
G
E
GX
T
G
F
X
EE
…
31
Conclusions
• If an agent’s beliefs are correct and complete then the game will evolve exactly as illustrated in the simulation
• The outcome of the game was affected from the instantiation of a new argument
• The strategic consideration here is for the proponent to avoid introducing arguments that could lead to the instantiation of new arguments which in turn might lead to an undesirable outcome
16/11/2011 King's College London Department of Informatics
32
Conclusions
• Though the soundness and fairness of dialogue systems that rely on abstract AFs can be shown for the purely abstract approach, we argue that such an approach is inadequate, as it fails to accommodate the fact that new arguments can be made available during the course of a dialogue
• The soundness of such systems is compromised
16/11/2011 King's College London Department of Informatics
33
Other characteristics of the system
• The opposing participants may also employ their different preference orderings on arguments, rules, or premises as those are described by ASPIC+
• The notion of attack in its three different forms is employed in the proposed system rather than that of defeat, thus we are treating preferences as moves in the dialogue
16/11/2011 King's College London Department of Informatics
34
Other characteristics of the system
16/11/2011 King's College London Department of Informatics
A
B
A>B
g → f
s → ¬ f
35
Future directions
16/11/2011 King's College London Department of Informatics
• Research strategising in iterative dialogues• Develop methodologies for building an opponent model
– Account for the possibility where a participant may be in error in its modeling, or;
– May hold beliefs about its opponent’s knowledge with varying degrees of certainty
King's College London Department of Informatics
3616/11/2011
Thanks