Asynchronous Interaction Patterns for Mining Multi-Agent System Models from Event Logs

Asynchronous Interaction Patterns for Mining Multi-Agent
System Models from Event Logs
Roman A. Nesterov1,2, Irina A. Lomazova1 
1 National Research University Higher School Of Economics (Moscow, Russia)
2 Università degli Studi di Milano-Bicocca (Milano, Italia)

View Slide

Outline
● Introduction
● Process Discovery for Multi-Agent Systems
● Service Interaction Patterns
● Modeling and Refining Service Interaction Patterns
● Experimental Evaluation
● Conclusion and Future Work
2 / 16

View Slide

Process Discovery
3 / 16

View Slide

Process Discovery
● Analysis of the observed behavior of information systems 
(ERP, BPM, HRM etc.).
● Construction of as-is (executable) process models from event logs.
3 / 16

View Slide

Discovering Process Models From Logs of MAS
● Interaction of several independent agents ➝ sophisticated behavior.
● Model structure ➝ behavior of agents and their interaction.
4 / 16

View Slide

t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
4 / 16
N

View Slide

t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
4 / 16
1. Simulate N and obtain an event log L.
2. Rediscover a model directly from L.
N

View Slide

t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
4 / 16
N N*

View Slide

t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
Two models of the 
same system
4 / 16
N*

View Slide

Compositional Process Discovery*
● System model is composed from agent models guided by a specification of
their asynchronous interaction.
● System model preserves properties of agent behavior.
● System model structure explicitly indicates interaction.
● Channel-composition of generalized workflow nets.
● Abstraction/refinement via -morphisms on Petri nets.
α
5 / 16
* Bernardinello L., Lomazova I., Nesterov R., Pomello L. Compositional Discovery
of Workflow Nets from Event Logs Using Morphisms (ATAED 2018).

View Slide

Service Interaction Patterns*
● A Systematic approach to the problem of modeling complex interactions.
● They describe typical ways of interactions.
● They describe agent interactions at the abstract level.
 
6 / 16
* Barros A. et al. Service Interaction Patterns (BPM 2005).
* Campagna D. Et al. BPMN 2.0 and the Service Interaction Patterns… (ICSOFT 2014).

View Slide

Service Interaction Patterns
● Bilateral/Multilateral interaction patterns.
● Single/Multiple transmission patterns.
6 / 16

View Slide

Service Interaction Patterns
● Bilateral/Multilateral interaction patterns;
● Single/Multiple transmission patterns.
6 / 16

View Slide

Bilateral Asynchronous Service Interaction Patterns
7 / 16

View Slide

Bilateral Asynchronous Service Interaction Patterns
Pattern Description
Send (Receive) SIP-1 An agent X sends (receives) a message to (from) an agent Y.
Concurrent Send (Receive) SIP-2
An agent X concurrently sends (receives) several messages (>1) to
(from) an agent Y.
Sending (Receiving) Choice SIP-3
An agent X sends (receives) exactly one out of two (or more)
alternative message sets to (from) an agent Y.
Send+Receive SIP-4
An agent X sends a message to an agent Y. Subsequently, Y sends a
response to X.
Concurrent Send+Receive SIP-5
An agent X concurrently sends several messages (>1) to an agent Y.
Then Y sends a response to each message received from X.
Sending+Receiving Choice SIP-6
An agent X sends exactly one out of two (or more) alternative message
sets to an agent Y . Subsequently, Y sends a corresponding response
to a message received from X.
Multiple Send+Receive SIP-7
The iterative implementation of SIP-4, s.t. message exchange process
continues till an Agent X does not need responses from an Agent Y.
7 / 16

View Slide

Generalized Workflow Nets
A GWF-net is a quintuple , where:
• is a Petri net,
• is an initial state and
• is a final state.
N = (P, T, F, m0
, mf
)
m0
⊆ P
(P, T, F)
mf
⊆ P
8 / 16

View Slide

• a bipartite graph,
• an (initial) state is indicated by putting
tokens inside places,
t1
t5
t2
t6
t4
t3
t8
t7
transition
place
initial state
ﬁnal state
8 / 16

View Slide

• a bipartite graph,
• an (initial) state is indicated by putting
tokens inside places,
• transition firings change a system state.
t1
t5
t2
t6
t4
t3
t8
t7
transition
place
initial state
ﬁnal state
8 / 16

View Slide

A GWF-net is a quintuple , where:
• is a Petri net,
• is an initial state and
• is a final state.
A GWF-net is sound if and only if:
• a final state is reachable from any reachable state (proper termination),
• whenever a final state is reached, no tokens are left in other places (clean
termination),
• every transition has the ability to fire (no "dead" transitions).
N = (P, T, F, m0
, mf
)
m0
⊆ P
(P, T, F)
mf
⊆ P
8 / 16

View Slide

GWF-nets for Bilateral Interaction Patterns
s r s1
s2
r1
r2
s1
r2
r1
s2
s11
s12
r21
r22
r11
r12
s21
s22
s11
s12
r21
r22
r11
r12
s21
s22
s12
r22
r11
s11
s21
r21
s1
s2
r1
r2
SIP-1
SIP-2 SIP-3 SIP-4
SIP-5 SIP-6
SIP-7
9 / 16

View Slide

Refining Send+Receive Pattern (SIP-4)
s1
r2
r1
s2
10 / 16

View Slide

s1
r2
r1
s2
N1
N2
10 / 16

View Slide

s1
r2
r1
s2
N1
N2
N1
⊕C
N2
10 / 16

View Slide

N1
⊕C
N2
s1
r1
t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
'
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
N1
N'1
φ: N′
1
→ N1
s1
r2
r1
s2
N1
N2
10 / 16

View Slide

N1
⊕C
N2
s1
r1
t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
'
AAACynicjVHLSsNAFD2Nr1pfVZdugkVwFRI3uiy4ceGign1AWyRJp+3QvJhMCqW48wfc6oeJf6B/4Z1xCmoRnZDkzLnn3Jl7b5BFPJeu+1qyVlbX1jfKm5Wt7Z3dver+QStPCxGyZphGqegEfs4inrCm5DJinUwwPw4i1g4mlyrenjKR8zS5lbOM9WN/lPAhD31JVLs39UU25nfVmuu4etnLwDOgBrMaafUFPQyQIkSBGAwJJOEIPnJ6uvDgIiOujzlxghDXcYZ7VMhbkIqRwid2Qt8R7bqGTWivcubaHdIpEb2CnDZOyJOSThBWp9k6XujMiv0t91znVHeb0T8wuWJiJcbE/uVbKP/rU7VIDHGha+BUU6YZVV1oshS6K+rm9peqJGXIiFN4QHFBONTORZ9t7cl17aq3vo6/aaVi1T402gLv6pY0YO/nOJdB68zxXMe7OavVHTPqMo5wjFOa5znquEIDTV3lI57wbF1bwppZ80+pVTKeQ3xb1sMHFVeSEQ==
N1
N'1
φ: N′
1
→ N1
s1
r2
r1
s2
N1
N2
t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
N'1 N'2
10 / 16

View Slide

N1
⊕C
N2
s1
r2
r1
s2
N1
N2
t1
t2
s1
t4
t3
r11
r12
t5
t6
t7
t8
t9
q1
q2
r21
r22
s21
s22
q4
q3
q5
q6
q7
q8
N'1 N'2
N1
⊕C
N2
sound
N′
1
⊕C
N′
2
sound
φi
: N′
i
→ Ni
10 / 16

View Slide

Experimental Evaluation: Layout
ABCDG
HIJKLNOP
QRSTUVW
XYZACDEH
IJKLMN…
System
event log
ABCILOR
TUXYAD
EJKLN…
DGHJKN
PQSVWZ
CHIM…
Agent 1
event log
Agent 2
event log
ﬁlter
B
…
… …
D G
… …
…
discover
D
B
Agent 2
behavior
Agent 1
behavior
Interface
D G
… …
…
B
…
… …
compose
Complete system model
SIPs
S W
… …
…
A
…
… …
System
simulate
instantiate
O G
… …
…
U
…
… …
N
…
…
discover
Directly discovered system model
Compare
quality
11 / 16

View Slide

Experimental Evaluation: Neighboring Transitions
Let be a GWF-net and .
Two transitions are called neighboring iff there is a path in
connecting them, where other transitions along this path are silent.
N = (P, T, F, m0
, mf
) λ : T → L ∪ {τ}
t1
, t2
(λ(ti
) ≠ τ) N
12 / 16

View Slide

Let , s.t. .
denotes the number of neighboring transition pairs in .
Symmetric pairs are counted as a single one.
N = (P, T, F, m0
, mf
) λ : T → L ∪ {τ}
t1
, t2
(λ(ti
) ≠ τ) N
T = T1
∪ T2
T1
∩ T2
= Ø
NbN
(T1
× T2
) ∪ (T2
× T1
)
12 / 16

View Slide

N = (P, T, F, m0
, mf
) λ : T → L ∪ {τ}
t1
, t2
(λ(ti
) ≠ τ) N
t1
t2
t3
q1
q2
q3
q4
t2
t1
q1
q2
q3
t3
12 / 16

View Slide

Experimental Evaluation: Instruments
• Agent model discovery and event log filtration: 
ProM and Inductive Miner
• Artificial event log generation: 
ProM plug-in GENA (extended with MAS simulation)*
In our experiments, artificial event logs contain 5000 traces.
13 / 16
* Nesterov R.A., Mitsyuk A.A., Lomazova I.A. Simulating Behavior of Multi-Agent Systems
with Acyclic Interactions of Agents // Proceedings of the ISP RAS. 2018.

View Slide

Experimental Evaluation: Results
SIP-1 SIP-2 SIP-3 SIP-4 SIP-5 SIP-6 SIP-7
Direct Discovery
Places 35 59 42 49 64 72 29
Transitions 41 60 66 46 69 84 34
Arcs 98 140 112 120 170 198 78
NbN 22 95 25 61 85 117 22
Precision 0,689 0,5646 0,7768 0,6733 0,4554 0,6586 0,6495
Compositional Discovery
Places 33 55 35 46 69 66 29
Transitions 32 49 35 39 57 72 28
Arcs 76 133 86 102 155 157 72
NbN 4 6 6 4 7 14 12
Precision 0,8017 0,4477 0,8162 0,6745 0,4342 0,8414 0,8500
14 / 16

View Slide

Experimental Evaluation: Results
• Compositionally discovered models are more compact with respect to the
number of nodes.
• Precision of composed models is mainly higher in comparison with directly
discovered models.
• The number of neighboring transition pairs are minimized to the number of
transition pairs involved in the asynchronous agent interaction.
14 / 16

View Slide

Conclusion and Future Work Directions
• Formal models for service interaction patterns (SIPs) — (channel-composed)
generalized workflow nets.
• The illustration to the formal compositional approach to process discovery —
refinement of abstract interaction patterns.
• The overall experimental results demonstrates the value of the compositional
approach.
• Multilateral interaction patterns,
• Systematic approach to constructing morphisms,
• Improvement of the metric based on the notion of neighboring transitions.
15 / 16

View Slide

Thank you for your
attention!
Roman A. Nesterov, 
[email protected]; [email protected]
16 / 16

View Slide

Asynchronous Interaction Patterns for Mining Multi-Agent System Models from Event Logs

Asynchronous Interaction Patterns for Mining Multi-Agent System Models from Event Logs

Exactpro
PRO

More Decks by Exactpro

Other Decks in Research

Featured

Transcript

Asynchronous Interaction Patterns for Mining Multi-Agent System Models from Event Logs

Asynchronous Interaction Patterns for Mining Multi-Agent System Models from Event Logs

Exactpro PRO

More Decks by Exactpro

Other Decks in Research

Featured

Transcript

Exactpro
PRO