An Automated Planning Approach for Generating Argument ... · An Automated Planning Approach for...

An Automated Planning Approach for GeneratingArgument Dialogue Strategies

Tanja Daub, Elizabeth Black and Amanda Coles

King’s College London

[email protected]

Cardiff Argumentation ForumJuly 7, 2016

1/24

2/24

Background Planning Argument Strategies Future Work

Overview

1 BackgroundPersuasion DialoguesClassical PlanningPlanning a DialoguePolicies

2 Planning Argument StrategiesSimple Strategies vs PoliciesGenerating a Policy from Simple Strategies

3 Future Work

Tanja Daub, Elizabeth Black and Amanda Coles KCL

An Automated Planning Approach for Generating Argument Dialogue Strategies

3/24


Persuasion Dialogues

Agents have conflicting views on a topic

Proponent’s goal: convince opponent toaccept the topic

Dialogue terminates when the opponentaccepts the topic or when neither agentasserts any more arguments

A

B

C

D

E



4/24


Argument Strategies

Existing work

AI Planning approach for simple persuasion dialogues[Black et al., 2014]

Mixed Observable Markov Decision Processes, assumesprobabilistic knowledge of opponent strategy[Hadoux et al., 2015]

Minimax algorithm [Rienstra et al., 2013]



5/24


Classical Planning

A classical planning problem consists of

a set of state variables

a set of actions defined by preconditions and effects

a start state

a goal state



6/24


Persuasion Dialogues as Planning Problems

[Black et al., 2014]

a set of state variables7→ different dialogue states

a set of actions defined by preconditions and effects7→ asserting arguments

a start state7→ initial knowledge of proponent and opponent

a goal state7→ topic is acceptable to the opponent



7/24


Planning a Dialogue

Opponent modelsM0 = {B}M1 = {C}M2 = {B,C}

Simple strategy{A,D}, {E}

A

B

C

D

E



8/24


Policies

A policy is a set of state-action-pairs that determines which actionshould be performed in which state



9/24


Simple Strategies vs Policies

A

B1B0 B2

C2C1C0

Opponent modelsM0 = {B0} : 0.3M1 = {B1} : 0.5M2 = {B2} : 0.2

Simple strategy{A,C1}, {C0}p = 0.8



10/24


Simple Strategies vs Policies

A

B1B0 B2

C2C1C0

Opponent modelsM0 = {B0} : 0.3M1 = {B1} : 0.5M2 = {B2} : 0.2

Policy(s0, aA)(sB0 , aC0)(sB1 , aC1)(sB2 , aC2)p = 1



11/24


Generating a Policy from Simple Strategies

A

B2B1B0 B3 B4

C2C0 C1 C3 C4



12/24


Finding a Simple Strategy

Opponent modelsM0 = {B0,B2} : 1/3M1 = {B2,B4} : 1/3M2 = {B1,B3} : 1/3

Simple strategy π0{A,C0,C2}, {C4}p = 2/3

A

B2B1B0 B3 B4

C2C0 C1 C3 C4



13/24


Generating a Policy

Opponent models

M0 = {B0,B2}M1 = {B2,B4}

}π0

M2 = {B1,B3}}

?

?

?

B0,B2,B4 B1,B3

π0 ?



14/24


Replanning for Failed Cases

Opponent modelsM0 = {B0,B2} : 1/3M1 = {B2,B4} : 1/3M2 = {B1,B3} : 1

Simple strategy π1{A,C1,C3}p = 1

Merge simple strategies into policy

A

B2B1B0 B3 B4

C2C0 C1 C3 C4



15/24


Merging Simple Strategies into a Policy

Opponent models

M0 = {B0,B2}M1 = {B2,B4}

}π0 = {A,C0,C2}, {C4}

M2 = {B1,B3}}π1 = {A,C1,C3}

Policy(s0, {aA})(sB0 , π0)(sB1 , π1)(sB2 , π0)(sB3 , π1)(sB4 , π0)p = 1

A

B0,B2,B4 B1,B3

π0 π1



16/24


Generating a Policy from Simple Strategies

A

B1B0 B2 B3

C0 C1 C2 C3



17/24


Finding a Simple Strategy

Opponent modelsM0 = {B0,B2} : 0.25M1 = {B0,B1} : 0.25M2 = {B1,B2} : 0.25M3 = {B1,B3} : 0.25

Simple strategy π0{A,C0,C2}, {C1}p = 0.75

A

B1B0 B2 B3

C0 C1 C2 C3



18/24


Generating a Policy

Opponent models

M0 = {B0,B2}M1 = {B0,B1}M2 = {B1,B2}

π0

M3 = {B1,B3}}

?

??

?

B0,B2 B3B1

π0 ??



19/24


Finding more Simple Strategies

Opponent modelsM0 = {B0,B2} : 0.25M1 = {B0,B1} : 1/3M2 = {B1,B2} : 1/3M3 = {B1,B3} : 1/3

Simple strategy π0{A,C0,C2}, {C1}p = 2/3

A

B1B0 B2 B3

C0 C1 C2 C3



20/24


Finding more Simple Strategies

Opponent modelsM0 = {B0,B2} : 0.25M1 = {B0,B1} : 1/3M2 = {B1,B2} : 1/3M3 = {B1,B3} : 1

Simple strategy π1{A,C1,C3}p = 1

A

B1B0 B2 B3

C0 C1 C2 C3



21/24



Opponent models

M0 = {B0,B2}M1 = {B0,B1}M2 = {B1,B2}

π0 = {A,C0,C2}, {C1}

M3 = {B1,B3}}π1 = {A,C1,C3}

A

B0,B2 B3B1

π0 π1?

B0 B2 B3



22/24



Opponent models

M1 = {B0,B1}}{C0,C2}

M2 = {B1,B2}}

{C2}

M3 = {B1,B3}}{C1,C3}

A

B0,B2 B3B1

π0 π1?

B0 B2 B3



23/24


Future Work

Implement this approach and perform experiments todetermine both its scalability and the quality of the policiescompared to the optimal

How can we identify problems where a policy would performbetter than a simple strategy?

What is the best simple strategy to start with?

How can we deal with more complex dialogue scenarios andopponent strategies?



24/24


References

E. Black, A. Coles, S. Bernardini (2014)

Automated planning of simple persuasion dialogues

Computational Logic in Multi-Agent Systems, LNCS vol. 8624, Springer,87 - 104.

E. Hadoux, A. Beynier, N. Maudet, P. Weng, A. Hunter (2015)

Optimization of probabilistic argumentation with Markov decision models.

Proceedings of the Twenty-Fourth International Joint Conference onArtifiial Intelligence, AAAI Press, 2004 - 2010.

T. Rienstra, M. Thimm, N. Oren (2013)

Opponent models with uncertainty for strategic argumentation.

Proceedings of the Twenty-Third International Joint Conference onArtficial Intelligence, AAAI Press, 332 - 338.



An Automated Planning Approach for Generating Argument ... · An Automated Planning Approach for...

Documents

Transcript of An Automated Planning Approach for Generating Argument ... · An Automated Planning Approach for...