Agreeing on Institutional Goals for Multi-Agent Societies D. Gaertner - - PowerPoint PPT Presentation

▶

Jul 09, 2023 375 likes •774 views

Agreeing on Institutional Goals for Multi-Agent Societies D. Gaertner 1 , 2 , J.-A. Rodriguez 2 , F. Toni 1 1 Department of Computing, Imperial College, London, United Kingdom 2 Artificial Intelligence Research Institute, IIIA-CSIC, Bellaterra,

SLIDE 1

Agreeing on Institutional Goals for Multi-Agent Societies

D. Gaertner1,2, J.-A. Rodriguez2, F. Toni1

1Department of Computing, Imperial College, London, United Kingdom 2Artificial Intelligence Research Institute, IIIA-CSIC, Bellaterra, Spain

SLIDE 2

Aims and Objectives

1 find agreement on a common set of goals 2 using tried-and-tested argumentation technology 3 sketch two approaches - central and distributed

SLIDE 3

Aims and Objectives

1 find agreement on a common set of goals 2 using tried-and-tested argumentation technology 3 sketch two approaches - central and distributed

SLIDE 4

Aims and Objectives

1 find agreement on a common set of goals 2 using tried-and-tested argumentation technology 3 sketch two approaches - central and distributed

SLIDE 5

Univ. of Texas paper

Combining Job and Team Selection Heuristics

Similarities: coalition formation large-scale, dynamic agent environment Differences: quantitative experiments vs. qualitative analysis builds on previous work - more mature and formal

SLIDE 6

King’s College paper

Argumentation Heuristic for Normative Conflicts

Similarities: use of argumentation technology to address COIN problems both needed for normative MAS - ours initially, theirs at run-time Differences: Dung’s Abstract Argumentation vs. Assumption-based Argumentation focus on individual agent vs. focus on collaboration

SLIDE 7

Argumentation Scenario Central Approach Distributed Approach

1 Assumption-based Argumentation

Theory Practice

2 Scenario

Goals Rules

3 Central Approach

Construction Scenario - Revisited

4 Distributed Approach

Construction Scenario - Revisited

5 Conclusions

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 8

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Formal Definition g

An assumption-based framework is a tuple L, R, A, where (L, R) is a deductive system A ⊆ L is the set of candidate assumptions if c ∈ A, then there exists no inference rule of the form c ← c1, . . . , cn ∈ R is a (total) mapping from A into L, where α is the contrary of α

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 9

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Example

Let L, R, A, be the assumption-based framework: L = {p, q, r, s, t, ¬p, ¬q, ¬r, ¬s, ¬t} R consists of p ← q, r q ← s ¬r ← t ¬t ← A = {r, s, t} r = ¬r, s = ¬s, t = ¬t {r, s} ⊢ p

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 10

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Semantics

A set of assumptions is admissible iff it does not attack itself and counter-attacks every set of assumptions attacking it complete iff it is admissible and contains all assumptions it can defend ground iff it is minimally complete ideal iff it is admissible and contained in all maximally admissible sets

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 11

Argumentation Scenario Central Approach Distributed Approach Theory Practice

CaSAPI g

system that determines acceptability and support of claims using assumption-based argumentation frameworks according to three semantics providing structured arguments and their inter-relationships

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 12

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Example in CaSAPI - Input

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 13

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Example in CaSAPI - Output

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 14

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Example in CaSAPI - Output

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 15

Argumentation Scenario Central Approach Distributed Approach Theory Practice

Example in CaSAPI - GUI

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 16

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - General

individual agents (A, B, C) in free and unregulated environments collaboration to fabricate motorbikes diverse goals:

generic (e.g. good communication, sustainability of collaboration) business (e.g. good profits for all participants) domain (e.g. produce X bikes, few delivery delays)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 17

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - Goals

g1: to produce 100 motorbikes this week g2: to always clear the assembly line at the end of each day g3: to produce 100 sidecars this week g4: to improve/foster relations between the three collaborators g5: to make the institution sustainable/repeatable g6: to make Carles as the leader of this collaboration

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 18

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - Rules

g1: to produce 100 motorbikes this week

achievable in two ways: assuming sufficiently many spare parts are in stock a 3rd party provider exists AND assuming outsourcing is acceptable g1 ← a1 g1 ← r1, c1

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 19

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - Rules

g1: to produce 100 motorbikes this week

achievable in two ways: assuming sufficiently many spare parts are in stock a 3rd party provider exists AND assuming outsourcing is acceptable g1 ← a1 g1 ← r1, c1

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 20

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - Rules

g1: to produce 100 motorbikes this week

achievable in two ways: assuming sufficiently many spare parts are in stock a 3rd party provider exists AND assuming outsourcing is acceptable g1 ← a1, x1 g1 ← r1, c1, x8

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 21

Argumentation Scenario Central Approach Distributed Approach Basics Goals Rules

Scenario - Rules

IKBA IKBB IKBC g1 ← a1, x1 g2 ← q1, x5 g1 ← r1, c1, x8 g2 ← p1, x2 g4 ← b1, q2, x6 g6 ← r2, x9 g3 ← p2, a2, x3 g5 ← q3, b2, x7 g3 ← a3, x4 p1 q1 ← b3 r3 p2 ← a1 q2 ¬a3 ← r3 q3 ¬b1 ← r4 ¬b2 r4 ← r5, c2 r5 ← c3 SKB contains ¬a1 and r1 ← r3

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 22

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Semantical Considerations

A goal is “acceptable” if there exists an acceptable set of assumptions according to a given semantics. credulous semantics won’t work

ut of the skeptical semantics we choose the ground one

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 23

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Formal Definition

For k agents, we construct the ABA framework as follows: R = SKB ∪ n

k=0 IKBk

A = n

k=0 Ak where Ak are the assumptions of agent k including its

applicability assumptions x′ = {y | y = xi and i is an agent }

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 24

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Scenario - Revisited

IKBA IKBB IKBC g1 ← a1, x1 g2 ← q1, x5 g1 ← r1, c1, x8 g2 ← p1, x2 g4 ← b1, q2, x6 g6 ← r2, x9 g3 ← p2, a2, x3 g5 ← q3, b2, x7 g3 ← a3, x4 p1 q1 ← b3 r3 p2 ← a1 q2 ¬a3 ← r3 q3 ¬b1 ← r4 ¬b2 r4 ← r5, c2 r5 ← c3 SKB contains ¬a1 and r1 ← r3

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 25

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: computationally straight-forward CON: general issues of centralised systems

performance issues / bottleneck / vulnerability to attacks / trust issues

CON: privacy

all knowledge must be shared (e.g. private business rules)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 26

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: computationally straight-forward CON: general issues of centralised systems

performance issues / bottleneck / vulnerability to attacks / trust issues

CON: privacy

all knowledge must be shared (e.g. private business rules)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 27

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: computationally straight-forward CON: general issues of centralised systems

performance issues / bottleneck / vulnerability to attacks / trust issues

CON: privacy

all knowledge must be shared (e.g. private business rules)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 28

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Outline

each agent runs CaSAPI engine communicate arguments compute attacks locally and communicate successful ones restrictions: shared notion of contrary + shared language

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 29

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Protocol Scheme

Agent (Initiator) broadcasts Agent Agent Agent goal as argument attack an argument

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 30

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Scenario - Revisited

IKBA IKBB IKBC g1 ← a1, x1 g2 ← q1, x5 g1 ← r1, c1, x8 g2 ← p1, x2 g4 ← b1, q2, x6 g6 ← r2, x9 g3 ← p2, a2, x3 g5 ← q3, b2, x7 g3 ← a3, x4 p1 q1 ← b3 r3 p2 ← a1 q2 ¬a3 ← r3 q3 ¬b1 ← r4 ¬b2 r4 ← r5, c2 r5 ← c3 SKB contains ¬a1 and r1 ← r3

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 31

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Protocol Instance

Betty broadcasts Adrian Carles g_4 as argument with support b_1 Adrian Betty broadcasts attack

n b_1 with

support {c_2,c_3} Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 32

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: less vulnerability + distributed workload PRO: privacy CON: communication overhead

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 33

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: less vulnerability + distributed workload PRO: privacy CON: communication overhead

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 34

Argumentation Scenario Central Approach Distributed Approach Construction Scenario - Revisited Discussion

Pros and Cons

PRO: less vulnerability + distributed workload PRO: privacy CON: communication overhead

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 35

Argumentation Scenario Central Approach Distributed Approach

Limitations and Future Work

(lessen the) honesty requirement relationship between individual and institutional goals favouritism / collusion investigate different semantics and preferences over goals trust and reputation agreeing about other institutional notions (e.g. norms, ...)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 36

Argumentation Scenario Central Approach Distributed Approach

Limitations and Future Work

(lessen the) honesty requirement relationship between individual and institutional goals favouritism / collusion investigate different semantics and preferences over goals trust and reputation agreeing about other institutional notions (e.g. norms, ...)

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 37

Argumentation Scenario Central Approach Distributed Approach

Conclusions

new approach to the problem of determining institutional goals first step in the direction of automating institution creation novel use of argumentation technology

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies

SLIDE 38

Argumentation Scenario Central Approach Distributed Approach

Thank you!

Questions ?

Gaertner et al. Agreeing on Institutional Goals for Multi-Agent Societies