05 MAS Distributed Constraint Optimization

Multiagent Systems
Distributed Constraint Optimization
Francesco Amigoni
Many of the following slides are taken from:
• the “lecture slides provided by authors”, Chapter 12, available
at:
http://www.the-mas-book.info/index.html
• the “Tutorial on Multi-agent Distributed Constrained
Optimization” at AAMAS2019, available at:
https://www2.isye.gatech.edu/~fferdinando3/cfp/AAMAS19/
Francesco Amigoni
Introduction
Distributed Constraint Reasoning
Applications and Exemplar Problems
Complete algorithms for DCOPs
Approximated Algorithms for DCOPs
Conclusions
Chapter 12: Distributed Constraint Handling and

Optimization
A. Farinelli1 M. Vinyals1 A. Rogers2 N.R. Jennings2
1
Computer Science Department
University of Verona, Italy
2
Agents, Interaction and Complexity Group
School of Electronics and Computer Science
University of Southampton,UK
Multi-Agent Systems. MIT Press, 2012 (2nd edition), edited by

Gerhard Weiss. http://www.the-mas-book.info
Chapter 12: Distributed Constraint Handling and Optimization

Introduction
Conclusions
Constraints
Pervade our everyday lives

Are usually perceived as elements that limit solutions to the
problems we face

Introduction
Conclusions
Constraints
From a computational point of view, they:

Reduce the space of possible solutions
Encode knowledge about the problem at hand
Are key components for efficiently solving hard problems

Introduction
Conclusions
Constraint Processing
Many different disciplines deal with hard computational problems that

can be made tractable by carefully considering the constraints that
define the structure of the problem.
Planning Operational Automated Reasoning Computer

Scheduling Research Decision Theory Vision

Introduction
Conclusions
Constraint Processing in Multi-Agent Systems
Focus on how constraint processing can be used to address

optimization problems in Multi-Agent Systems (MAS) where:
A set of agents must come to some agreement, typically via some

form of negotiation, about which action each agent should take in
order to jointly obtain the best solution for the whole system.
M2
M1
A1 A2 A2 A3

Introduction
Conclusions
Distributed Constraint Optimization Problems (DCOPs)

We will consider Distributed Constraint Optimization Problems (DCOP)
where:
Each agent negotiates locally with just a subset of other agents

(usually called neighbors) that are those that can directly influence
his/her behavior.
M2
M1
A2
A1 A3

Introduction
Conclusions
Constraint Networks
A constraint network N is formally defined as a tuple hX , D , C i where:
X = {x1 , . . . , xn } is a set of discrete variables;

D = {D1 , . . . , Dn } is a set of variable domains, which enumerate
all possible values of the corresponding variables; and
C = {C1 , . . . , Cm } is a set of constraints; where a constraint Ci is
defined on a subset of variables Si ✓ X which comprise the
scope of the constraint
r = |Si | is the arity of the constraint
Two types: hard or soft

Introduction
Conclusions
Hard constraints
A hard constraint Cih is a relation Ri that enumerates all the valid

joint assignments of all variables in the scope of the constraint.
Ri ✓ Di1 ⇥ . . . ⇥ Dir
Dj = Dk = {0, 1} Ri xj xk
0 1
1 0

Introduction
Conclusions
Soft constraints
A soft constraint Cis is a function Fi that maps every possible joint

assignment of all variables in the scope to a real value.
Fi : Di1 ⇥ . . . ⇥ Dir ! ¬
Fi xj xk
2 0 0
0 0 1
0 1 0
1 1 1

Introduction
Conclusions
Binary Constraint Networks
x1
Binary constraint networks are those where:
F 1,3
,2
Each constraint (soft or hard) is defined
F1
over two variables.
Every constraint network can be mapped to
a binary constraint network x2 x3
,4
requires the addition of variables and
F1
constraints
4
F2,
may add complexity to the model
They can be represented by a constraint
graph x4

Introduction
Conclusions
Different objectives, different problems
Constraint Satisfaction Problem (CSP)
Objective: find an assignment for all the variables in the network

that satisfies all constraints.
Constraint Optimization Problem (COP)
Objective: find an assignment for all the variables in the network

that satisfies all constraints and optimizes a global function.
Global function = aggregation (typically sum) of local functions.

F (x ) = Âi Fi (xi )

Introduction
Conclusions
A1
When operating in a
decentralized context:
a set of agents control
variables
A2
agents interact to find a A4
solution to the constraint
network
A3

Introduction
Conclusions
Two types of decentralized problems:

distributed CSP (DCSP)
distributed COP (DCOP)
Here, we focus on DCOPs.

Introduction
Conclusions
Distributed Constraint Optimization Problem (DCOP)
A DCOP consists of a constraint network N = hX , D , C i and a set of

agents A = {A1 , . . . , Ak } where each agent:
controls a subset of the variables Xi ✓ X
is only aware of constraints that involve variable it controls
communicates only with its neighbours

Introduction
Conclusions
Distributed Constraint Optimization Problem (DCOP)
Agents are assumed to be fully cooperative

Goal: find the assignment that optimizes the global function, not
their local local utilities.
Solving a COP is NP-Hard and DCOP is as hard as COP.

Introduction
Conclusions
Motivation
Why distribute?
Privacy
Robustness
Scalability

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Real World Applications
Many standard benchmark problems in computer science can be

modeled using the DCOP framework:
graph coloring
As can many real world applications:
human-agent organizations (e.g. meeting scheduling)
sensor networks and robotics (e.g. target tracking)

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Graph coloring
Popular benchmark
Simple formulation
Complexity controlled with few parameters:
Number of available colors
Number of nodes
Density (#nodes/#constraints)
Many versions of the problem:
CSP, MaxCSP, COP

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Graph coloring - CSP
Nodes can take k colors

Any two adjacent nodes should have different colors
If it happens this is a conflict
Yes! No!

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Graph coloring - Max-CSP
Minimize the number of conflicts
0 -1

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Graph coloring - COP
Different weights to violated constraints

Preferences for different colors
0 -2
-1
-3 -2
-1

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Graph coloring - DCOP
Each node:
controlled by one agent
Each agent:
Preferences for different colors
Communicates with its direct neighbours in the graph
-1
A1 A3 A1 and A2 exchange
-3 -2 preferences and conflicts
A3 and A4 do not
A2 A4 communicate
-1
Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Meeting Scheduling
Motivation:
Privacy
Robustness
Scalability

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Meeting Scheduling
In large organizations many people, possibly working in different

departments, are involved in a number of work meetings.

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Meeting Scheduling
People might have various private preferences on meeting start times
Better after 12:00am

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Meeting Scheduling
Two meetings that share a participant cannot overlap
Window: 15:00-18:00
Duration: 2h
Window: 15:00-17:00
Duration: 1h

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
DCOP formalization for the meeting scheduling problem
A set of agents representing participants

A set of variables representing meeting starting times according
to a participant.
Hard Constraints:
Starting meeting times across different agents are equal
Meetings for the same agent are non-overlapping.
Soft Constraints:
Represent agent preferences on meeting starting times.
Objective: find a valid schedule for the meeting while maximizing the
sum of individuals’ preferences.

Introduction
Graph coloring
Meeting Scheduling
Target Tracking
Conclusions
Target Tracking
Motivation:
Privacy
Robustness
Scalability

MOTIVATING DOMAIN:  
SENSOR NETWORK
x1 x2
x3 x4
x5 x6
AAMAS-19 Tutorials Fioretto, Picard, Tran-Thanh, Rust

5
DCOP
DISTRIBUTED CONSTRAINT OPTIMIZATION
a1 a2
x1 x2
a3 c2 a4
c1
x3 x4
a5 a6
x5 x6

26
Introduction
Applications and Exemplar Problems Search Based: ADOPT
Complete algorithms for DCOPs Dynamic Programming DPOP
Conclusions
Complete Algorithms
Always find an optimal solution

Exhibit an exponentially increasing coordination overhead
Very limited scalability on general problems.

Introduction
Conclusions
Complete Algorithms
Completely decentralised
Search-based.
Synchronous: SyncBB, AND/OR search
Asynchronous: ADOPT, NCBB and AFB.
Dynamic programming.
Partially decentralised
OptAPO
Next, we focus on completely decentralised algorithms

Introduction
Conclusions
Decentralised Complete Algorithms
Search-based Dynamic programming

Uses distributed search Uses distributed inference
Exchange individual values Exchange constraints
Small messages but Few messages but
. . . exponentially many . . . exponentially large
Representative: ADOPT [Modi Representative: DPOP [Petcu
et al., 2005] and Faltings, 2005]
We will discuss DPOP in detail

Introduction
Conclusions
DPOP
DPOP (Dynamic Programming Optimization Protocol) [Petcu and

Faltings, 2005]:
Based on the dynamic programming paradigm.
Special case of Bucket Tree Elimination Algorithm (BTE)
[Dechter, 2003].

Introduction
Conclusions
DFS arrangement
The same happens

for DPOP!
Before executing ADOPT, agents must be arranged in a depth

first search (DFS) tree.
DFS trees have been frequently used in optimization because
they have two interesting properties:
Agents in different branches of the tree do not share any
constraints;
Every constraint network admits a DFS tree.

Introduction
Conclusions
DPOP by example
x1 x1
Fi ,j xi xj
F 1,3
,2
F1
P2 , child !
2 0 0
, PP4 !
PP 3 , p
0 0 1
,4
F2,3
F1
x2 x3 0 1 0
ochild
seudo
1 1 1
pseud
x2
child
,4
F2
!
=>
P3
P4
, ch
ld ,
chi
x4
ild
DFS arrangement
!
Objective: find assignment x4 x3
with maximal value

Introduction
Conclusions
DPOP phases
Given a DFS tree structure, DPOP runs in two phases:

Util propagation: agents exchange util messages up the tree.
Aim: aggregate all info so that root agent can choose optimal
value
Value propagation: agents exchange value messages down the
tree.
Aim: propagate info so that all agents can make their choice given
choices of ancestors

Introduction
Conclusions
Sepi : set of agents preceding Ai in the pseudo-tree order that are

connected with Ai or with a descendant of Ai .
x1 Sep1 = 0/
P2 , child !
, PP4 !
PP 3 , p
ochild
seudo
pseud
x2 Sep2 = {x1 }
child
!
!
P3
P4
, ch
ld ,
chi
ild
!
Sep4 = {x1 , x2 } x4 x3 Sep3 = {x1 , x2 }

Introduction
Conclusions
Util message
The Util message Ui !j that agent Ai sends to its parent Aj can be

computed as:
!
N N
Ui !j (Sepi ) = max Uk !i ⌦ Fi , p
xi Ak 2Ci Ap 2Pi [PPi
Size exponential All incoming messages Shared constraints with

in Sepi from children parents/pseudoparents
The ⌦ operator is a join operator that sums up functions with different
but overlapping scores consistently.

Introduction
Conclusions
Join operator
F2,4 x2 x4 max{x4 } (F1,4 ⌦ F2,4 )

F1,4 ⌦ F2,4
2 0 0
x1 x2 x4 x1 x2 x4
0 0 1
4 0 0 0 0 0 0
0 1 0 max(4,0)
0 0 0 1 0 0 1
1 1 1
2 0 1 0
Project 0 1 0
Add max(2,1)
1 0 1 1 out x4 0 1 1
F1,4 x1 x4
2 1 0 0 1 0 0
2 0 0 max(2,2)
2 1 0 1 1 0 1
0 0 1
0 1 1 0 1 1 0
0 1 0 max(0,2)
2 1 1 1 1 1 1
1 1 1

Introduction
Conclusions
Complexity exponential to the largest Sepi .

Largest Sepi = induced width of the DFS tree ordering used.
A1 (root)
U2!1 x1 Sep2
U1!2
0 10
Sep4 1 5
A2 max(U3!2 ⌦ U4!2 ⌦ F1,2 )
x2
U4!2 x1 x2 U3!2 x1 x2 Sep3
U3
4 0 0 4 0 0
!2
4!
!
U
2 0 1 2
2 0 1
2 1 0 2 1 0
2 1 1 A3 2 1 1
A4
1 max(F1,3 ⌦ F2,3 )
2
max(F1,4 ⌦ F2,4 )
x4 x3
Introduction
Conclusions
Value message
Keeping fixed the value of parent/pseudoparents, finds the value that

maximizes the computed cost function in the util phase:
!
xi⇤ = arg max
xi
Â Uj !i (xi , xp⇤ ) + Â Fi ,j (xi , xj⇤ )
Aj 2Ci Aj 2Pi [PPi
S
where xp⇤ = Aj 2Pi [PPi {xj⇤ } is the set of optimal values for Ai ’s parent
and pseudoparents received from Ai ’s parent.
Propagates this value through children down the tree:
[
⇤
Vi !j = {xi = xi } [ {xs = xs⇤ }
xs 2Sepi \Sepj

Introduction
Conclusions
A1 x1⇤ = max U1!2 (x1 )

x1
V1!2
!
A2 x2⇤ = max(U3!2 (x1⇤ , x2 ) ⌦ U4!2 (x1⇤ , x2 ) ⌦ F1,2 (x1⇤ , x2 ))
x2
V2 !
4
2!
!
V
A4 A3
x4⇤ = max(F1,4 (x1⇤ , x4 ) ⌦ F2,4 (x2⇤ , x4 )) x3⇤ = max(F1,3 (x1⇤ , x3 ) ⌦ F2,3 (x2⇤ , x3 ))
x4 x3

Introduction
Conclusions
DPOP extensions
MB-DPOP [Petcu and Faltings, 2007] trades-off message size

against the number of messages.
A-DPOP trades-off message size against solution quality [Petcu
and Faltings, 2005(2)].

Example run of DPOP
SBB
B { } xi xj Cost  Cost  Cost  Cost 
(A,B) (A,C) (B,C) (B,D)
D 5 5 5 3
A { } 8 10 4 8
{ } C 20 20 3 10
{ } 3 3 3 3
Katsutoshi Hirayama, Makoto Yokoo: Distributed Partial Constraint Satisfaction Problem. CP 1997: 222-236

38
Francesco Amigoni
PSEUDO-TREE
B { } A { }
D
A { } B { }
{ } C
{ }
Definition: A spanning tree of the constraint C D
graph such that no two nodes in sibling subtrees { } { }

share a constraint in the constraint graph
60
DPOP
Pseudo-tree Ordering
A { }
B D (B,D)
r r 3
r g 8
g r 10 B { }
g g 3
UTIL C D
{ } { }
DPOP
A { }
B D (B,D)
r r 3
r g 8
min{3, 8} = 3
g r 10 B { }
g g 3
min{10, 3} = 3
MSG to B
B cost
UTIL C D
r 3
g 3 { } { }
DPOP
A B C (B,C) (A,C)
r r r 5 5 10 A { }
r r g 4 8 12
r g r 3 5 8
r g g 3 8 11
g r r 5 10 15
g r g 4 3 7 B { }
g g r 3 10 13
g g g 3 3 6
UTIL C D
{ } { }
DPOP
A B C (B,C) (A,C)
r r r 5 5 10 A { }
r r g 4 8 12
r g r 3 5 8
r g g 3 8 11
g r r 5 10 15
g r g 4 3 7 B { }
g g r 3 10 13
g g g 3 3 6
UTIL
MSG to B A
r
B
r 10
C D
r
g
g
r
8
7
{ } { }
AAMAS-19 Tutorials g g 6 Fioretto, Picard, Tran-Thanh, Rust
DPOP
A { }
Util Util
A B (A,B)
C D
r r 5 10 3 18
r g 8 8 3 19
g r 20 7 3 30
B { }
g g 3 6 3 12
UTIL C D
{ } { }
DPOP
A { }
Util Util
A B (A,B)
C D
r r 5 10 3 18
r g 8 8 3 19
g r 20 7 3 30
B { }
g g 3 6 3 12
MSG to A
A cost
UTIL C D
r 18
g 12 { } { }
DPOP
A { }
A cost
r 18
g 12
optimal cost = 12 B { }
UTIL C D
{ } { }
DPOP
A { }
A cost
r 18 VALUE
g 12
•Select value for A = ‘g’ B { }

•Send MSG A = ‘g’ to  
agents B and C
C D
Alternative version in which
value messages xi = xi* are { } { }
sent to children and
AAMAS-19 Tutorials
pseudo-children Fioretto, Picard, Tran-Thanh, Rust
DPOP
A { }
Util Util
A B (A,B)
C D
VALUE
r r 5 10 3 18
r g 8 8 3 19
g r 20 7 3 30
B { }
g g 3 6 3 12
•Select value for B = ‘g’

•Send MSG B = ‘g’ to   C D
agents C and D { } { }
DPOP
A B C (B,C) (A,C) Pseudo-tree Ordering
r r r 5 5 10
A { }
r r g 4 8 12
r g r 3 5 8 VALUE
r g g 3 8 11
g r r 5 10 15
g r g 4 3 7
g g r 3 10 13 B { }
g g g 3 3 6
•Select value for C = ‘g’ C D
{ } { }
DPOP
A { }
VALUE
B D (B,D)
r r 3
r g 8
g r 10 B { }
g g 3
•Select value for D = ‘g’ C D
{ } { }
Introduction
Local greedy methods: DSA-1, MGM-1 (Heuristic)
GDL-based approaches: Max-Sum (Heuristic)
Quality guarantees: k-optimality, region optimality, bounded Max-Sum
Conclusions
Why Approximate Algorithms
“Very often optimality in practical applications is not achievable”
Approximate algorithms
Sacrify optimality in favor of computational and communication
efficiency
Well-suited for large scale distributed applications:
sensor networks
mobile robots

Introduction
Conclusions
Centralized Local Greedy approaches
Start from a random assignment for all the variables

Do local moves if the new assignment improves the value (local
gain)
Local: changing the value of a small set of variables (in most
case just one)
The search stops when there is no local move that provides a
positive gain, i.e., when the process reaches a local maximum.

Introduction
Conclusions
Distributed Local Greedy approaches
When operating in a decentralized context:

Problem: Out-of-date local knowledge
Assumption that other agents do not change their values
A greedy local move might be harmful/useless
Solution:
Stochasticity on the decision to perform a move (DSA)
Coordination among neighbours on who is the agent that should
move (MGM)

Introduction
Conclusions
Distributed Stochastic Algorithm (DSA)
Activation probability to mitigate issues with parallel executions
[S. Fitzpatrick and L. Meetrens, 2003]
Initialize agents with a random assignment and communicate

values to neighbors
Each agent:
Generates a random number and executes only if it is less than
the activation probability
When executing choose a value for the variable such that the local
gain is maximized
Communicate and receive possible variables change to/from
neighbors

Example run of DSA
DSA ALGORITHM
Utility  Utility 
xi xj
(A,B) (B,C)
5 5
A B C 5 0
{ } { } { } 0 0
8 8

80
Francesco Amigoni
Example run of DSA
DSA ALGORITHM
Assume that all agents are activated
xi xj Utility  Utility 
U=10, Δ=5 (A,B) (B,C)
U=5, Δ=0 5 5
U=5 U=0
A B C 5 0
U=5, Δ=0 U=0, Δ=0 0 0
U=8, Δ=3 U=8, Δ=8 8 8

81
Francesco Amigoni
Example run of DSA
DSA ALGORITHM
Assume that only agent B is activated
Utility  Utility 
U=0, Δ=0 xi xj
(A,B) (B,C)
U=16, Δ=16 5 5
U=0 U=0
A B C 5 0
U=5, Δ=5 U=5, Δ=5 0 0
U=0, Δ=0 U=0, Δ=0 8 8

83
Francesco Amigoni
Example run of DSA
DSA ALGORITHM
Assume that all agents are activated
xi xj Utility  Utility 
U=0, Δ=-16 (A,B) (B,C)
U=16, Δ=0 5 5
U=8 U=8
A B C 5 0
U=0, Δ=-8 U=0, Δ=-8 0 0
U=8, Δ=0 U=8, Δ=0 8 8

84
Francesco Amigoni
Example run of DSA
DSA ALGORITHM
U=5 U=0
A B C Utility  Utility 
xi xj
(A,B) (B,C)
U=8 U=0 5 5
A B C
5 0
U=0 U=0
A B C 0 0
U=8 U=8 8 8
A B C
One possible execution trace

85
Francesco Amigoni
Introduction
Conclusions
DSA-1: discussion
Extremely low computation/communication

Shows an anytime property (not guaranteed)
Activation probability:
Must be tuned
Domain dependent (no general rule)

Introduction
Conclusions
Maximum Gain Message (MGM-1)
Coordination among neighbours to decide which single agent can

perform the move.
[R. T. Maheswaran et al., 2004]
Initialize agents with a random assignment and communicate

values to neighbors
Each agent:
Compute and exchange possible gains
Agent with maximum (positive) gain executes
Communicate and receive possible variables changes to/from
neighbors

Example run of MGM-1
Domains: {green, blue, red}

Constraints: ≠
1 2
3 4
5 6
Satisfied constraints cost 0, while

violated constraints cost 1
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
❷❸❹ ❶❸❹❺
1 2
❶❷❺❻ ❶❷❺❻
3 4
❷❸❹❻ ❸❹❺
5 6
Francesco Amigoni
0 ❶ 0 ❷
1 2
0 ❸ 1 ❹
3 4
1 ❺ 2 ❻
5 6
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
0 ❶ 0 ❷
1 2
0,0,1 0,0,1,1
0 ❸ 1 ❹
3 4
0,0,1,2 0,0,1,2
1 ❺ 2 ❻
5 6
0,0,1,2 0,1,1
Francesco Amigoni
0 ❶ 0 ❷
1 2
0,0,1 0,0,1,1
0 ❸ 1 ❹
3 4
0,0,1,2 0,0,1,2
1 ❺ 2 ❻
5 6
0,0,1,2 0,1,1
Francesco Amigoni
0 ❶ 0 ❷
1 2
0,0,1 0,0,1,1
0 ❸ 1 ❹
3 4
0,0,1,2 0,0,1,2
1 ❺ 2 ❻
5 6
0,0,1,2 0,1,1
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
❷❸❹ ❶❸❹❺
1 2
❶❷❺❻ ❶❷❺❻
3 4
❷❸❹❻ ❸❹❺
5 6
Francesco Amigoni
0 ❶ 0 ❷
1 2
0 ❸ 0 ❹
3 4
0 ❺ 0 ❻
5 6
Francesco Amigoni
0 ❶ 0 ❷
1 2
0 ❸ 0 ❹
3 4
0 ❺ 0 ❻
5 6
Francesco Amigoni
Why?
Local maximum vs. global maximum

MGM has reached a local maximum and, changing only a value, cannot escape from it
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
1 2
3 4
5 6
Francesco Amigoni
Introduction
Conclusions
MGM-1: discussion
= More communication than DSA but still linear

= Empirically similar to DSA
Guaranteed to be anytime
Does not require any parameter tuning.

MGM vs. DSA
1 2
𝑈(𝑟𝑒𝑑, 𝑟𝑒𝑑) = 0
𝑈(𝑟𝑒𝑑, 𝑔𝑟𝑒𝑒𝑛) = 𝑈(𝑔𝑟𝑒𝑒𝑛, 𝑟𝑒𝑑) = 1
𝑈(𝑔𝑟𝑒𝑒𝑛, 𝑔𝑟𝑒𝑒𝑛) = – 1000
If both agents are allowed to alter their value in the same round, we could end up in the adverse
state (green, green)
When using DSA, there is always a positive probability for any time horizon that (green, green)
will be the resulting assignment
Francesco Amigoni
MGM vs. DSA
When applying MGM, the global utility strictly increases at each round until convergence
Francesco Amigoni
Introduction
Conclusions
Decentralised greedy approaches
Very little memory and computation

Anytime behaviours
Could result in very bad solutions
local maxima arbitrarily far from optimal

Introduction
Conclusions
The Max-Sum algorithm
Agents iteratively exchange messages to build a local function that

depends only on the variables they control
A1
x1
m1
1)
!
1 (x
!
F 1,3
!1 )
,2
m3
2)
3 !
2!
(x 3
x
F1
2 (x
1(
!
m
)
)
1
4!
,4
1!
(x 1
x4
F2,3
F1
4(
m
m2!3 (x3 )
)
x2 x3 =>
1!
!
m
A2 A3
,4
m3!2 (x2 )
!)
x2
F2
2(
4!
)
x4
m
4(
x4 m
2!
A4

Introduction
Conclusions
Max-Sum messages
At each execution step, each agent Ai sends to each of its neighbors

Aj the message:
!
mi !j (xj ) = aij + maxxi Fij (xi , xj ) + Â mk !i (xi )
k 2N (i )\j
Shared constraint with Aj All incoming messages except from Aj
where:
aij is a normalization constant added to all components of the
message so that Âxj mi !j (xj ) = 0
N (i ) is the set of indices for variables that are connected to xi

Introduction
Conclusions
Max-Sum by example
A1
2)
2 (x
1!
m
A2 A3
m3!2 (x2 )
!2 )
x
2(
4!
m
m1!2 (x2 ) = maxx1 (F1,2 (x1 , x2 ) + m3!1 (x1 ) + m4!1 (x1 ))
A4
Shared constraint with A2 All incoming messages except from A2

Introduction
Conclusions
Max-Sum assignments
At each iteration, each agent Ai :

computes its local function as:
zi (xi ) = Â mk !i (xi )
k 2N (i )
All incoming messages
sets its assignment as the value that maximizes its local function:
x̃i = arg max zi (xi )

xi

Introduction
Conclusions
Max-Sum by example
A1
2)
2 (x
1!
m
A2 A3
m3!2 (x2 )
!2 )
x
2(
4!
m
z1 (x1 ) = m2!1 (x1 ) + m3!1 (x1 ) + m4!1 (x1 )
A4
All incoming messages: from A4 , A3 and A1

Example run of Max-Sum
Consider this grid environment, divided in three areas (denoted a b c

by patterns)
Each area is controlled by an agent, which has to place a sensor A
in the area with the constraint that there cannot be two sensors
on the same row (A, B, C) nor on the same column (a,b,c)
B
Agents are {1,2,3}, their variables are the positions of the

sensors in their areas {P1, P2, P3}, their domains are {Aa, Ab}, C
{Ac, Bc, Cc}, and {Ba, Bb}, respectively, and the constraints are:
F12 P1 P2 F13 P1 P3 F23 P2 P3

-¥ Aa Ac -¥ Aa Ba 0 Ac Ba
0 Aa Bc 0 Aa Bb 0 Ac Bb
0 Aa Cc 0 Ab Ba -¥ Bc Ba
-¥ Ab Ac -¥ Ab Bb -¥ Bc Bb
0 Ab Bc 0 Cc Ba
0 Ab Cc 0 Cc Bb
Francesco Amigoni
Assume normalization constants a equal to 0

In the first execution round, agent 1 sends the following messages:
m1à2(P2) = a12 + maxP1 F12(P1,P2) = PAc2 m -¥
1à2
Bc 0
Cc 0
m1à3(P3) = a13 + maxP1 F13(P1,P3) = P3 m1à3

Ba 0
Bb 0
Similarly, agent 2 sends the following messages: P1 m2à1 P3 m2à3

Aa 0 Ba 0
Ab 0 Bb 0
And agent 3 sends the following messages: P1 m3à1 P2 m3à2

Aa 0 Ac 0
Ab 0 Bc -¥
Cc 0
Francesco Amigoni
Agent 1 computes z1(P1) = m2à1(P1) + m3à1(P1) = P1 z1

Aa 0
Ab 0
and selects one value that maximizes z1(P1), namely ~P1 = Aa (according to the lexicographic
order)
Agent 2 does the same: computes z2(P2) = m1à2(P2) + m3à2(P2) = P2 z2

Ac -¥
Bc -¥
Cc 0
and selects the value ~P2 = Cc
Agent 3 does the same: computes z3(P3) = m1à3(P3) + m2à3(P3) = P3 z3

Ba 0
Bb 0
and selects the value ~P3 = Ba (according to the lexicographic order)
Then, Max-Sum continues with the next iteration…
Francesco Amigoni
Introduction
Conclusions
Max-Sum on acyclic graphs
A1
Optimal on acyclic graphs
Different branches are independent
z functions provide correct estimations of A2 A3
agents contributions to the global problem
Convergence guaranteed in a polynomial
number of cycles
A4

Introduction
Conclusions
Max-Sum on cyclic graphs
On cyclic graphs, limited theoretical results:

Lack of convergence guarantees
When converges, it does converge to a neighborhood maximum
Neighborhood maximum: guaranteed to be greater than all other
maxima within a particular region of the search space

Introduction
Conclusions
Quality guarantees
So far, algorithms presented (DSA-1, MGM-1, Max-Sum) do not

provide any guarantee on the quality of their solutions
Quality highly dependent on many factors which cannot always
be properly assessed before deploying the system.
Particularly adverse behaviour on specific pathological instances.
Challenge:
Quality assessment on approximate algorithms

Introduction
Conclusions
Quality guarantees for approx. techniques
Key area of research

Address trade-off between guarantees and computational effort
Particularly important for:
Dynamic settings
Severe constrained resources (e.g. embedded devices)
Safety critical applications (e.g. search and rescue)

DCOP APPLICATIONS
• Scheduling Problems
• Taking DCOP to the Real World: Efficient Complete Solutions for Distributed Multi-Event Scheduling. AAMAS 2004
• Radio Frequency Allocation Problems
• Improving DPOP with Branch Consistency for Solving Distributed Constraint Optimization Problems. CP 2014
• Sensor Networks
• Preprocessing techniques for accelerating the DCOP algorithm ADOPT. AAMAS 2005
• Home Automation
• A Multiagent System Approach to Scheduling Devices in Smart Homes. AAMAS 2017, IJCAI 2016
• Traffic Light Synchronization
• Evaluating the performance of DCOP algorithms in a real world, dynamic problem. AAMAS 2008
• Disaster Evacuation
• Disaster Evacuation Support. AAAI 2007; JAIR 2017
• Combinatorial Auction Winner Determination
• H-DPOP: Using Hard Constraints for Search Space Pruning in DCOP. AAAI 2008

107
TRAFFIC FLOW CONTROL
• Given a set of traffic lights in adjacent intersections
• How coordinate them to create green waves?
N/S N/S N/S

W/E W/E W/E
de Oliveira, D., Bazzan, A. L., & Lesser, V.. Using cooperative mediation to coordinate traffic lights: a case study. AAMAS 2005: 371-378
Junges, R., & Bazzan, A. L. Evaluating the performance of DCOP algorithms in a real world, dynamic problem. AAMAS 2008: 463-470

110
• Agents: Each traffic light
• Values: Flow traffic direction
x1 x2 x3
N/S N/S N/S

W/E W/E W/E

111
• Agents: Each traffic light
• Values: Flow traffic direction
• Conflict if 2 neighboring signals choose different directions
x1 x2 x3
N/S N/S N/S

W/E W/E W/E

112
• Cost functions model the number of incoming vehicles  
• Maximize the traffic flow
x1 x2 x3
N/S N/S N/S

W/E W/E W/E

113
SMART DEVICES

114
HOME ASSISTANTS

115
SMART HOMES  
DEVICE SCHEDULING
battery charge
sensor
A smart home has:
• Smart devices (roomba, HVAC)
that it can control
• Sensors (cleanliness,
temperature)
• A set of locations
thermostat cleanliness
sensor
Ferdinando Fioretto, William Yeoh, Enrico Pontelli. "A Multiagent System Approach to Scheduling Devices in Smart Homes". AAMAS, 2017.
117
SMART HOMES  
DEVICE SCHEDULING
Smart device:
• A set of actions it can
perform (clean, charge)
• Power consumption
associated to each action.
Example 2 The scheduling rule (1) describes an ASR
  a goal state where the living room floor is at least
defining
75%Scheduling
clean (i.e., at least Rules:
75% of the floor is cleaned by a
vacuum cleaning robot) by 1800 hrs:
living room cleanliness 75 before 1800 (1)
and scheduling rules (2) and (3) describe PSRs stating that
the battery charge of the vacuum cleaning robot zv needs to
be between 0 and 100 % of its full charge at all the times:
AAMAS-19 Tutorials
zv battery charge Fioretto,
0 always (2)
Picard, Tran-Thanh, Rust
118 zv battery charge  100 always (3)
SMART HOMES   Example 2 The scheduling rule (1) describes an ASR
8
DEVICE SCHEDULING
> min(20, 100 !)
>
<
defining a goal state where the living room floor is at least 25
75% clean (i.e., at least 75% of the floor is cleaned by a > 0
vacuum cleaning robot) by 1800 hrs: >
:
?
In other words, at each tim
and scheduling rules (2) and (3) describe PSRs stating that will increase by 20% if it is
S the battery
R Rcharge
C of the
C vacuum
R R C robot
cleaning R zvDevice
needs to
Schedule decrease by 25% if it is ru
be between 0 and 100 % of its full charge at all the times: charge, and no change if it
0 15 30 30 30 45 60 60 75 Cleanliness (%)
zv battery charge 0 always (2) Example 4 Consider the

65 40 15 35 55 30 5 25 0 Battery Charge (%) room, where the only actu
zv battery charge  100 always (3) a vacuum cleaning robot z
[t !t ] sume the room is 0% clean
Goal 75 For a home hi , we denote with Rp a b a scheduling rule 75
robot zv is ⇠zt v at that time
60 over a state property p 2 PH [PZ , and time interval [ta , t60 b ].
Battery Charge (%)

Cleanliness (%)
Each scheduling rule indicates a goal state (either a desired t+1

cleanliness (0, ⇠zt v ) = 0 +
45
state of a home if it is an ASR or a required state of a de- 45 ⇢
min
30 vice or a home if it is a PSR) at a location `Rp 2 Li [ 30 Zi t
cleanliness (!, ⇠zv ) =
of a particular state property p that must hold over the time 0
15
interval [ta , tb ] ✓ T. Each rule is associated with a set 15of
0 actuators p ✓ Ai that can be used to reach the goal state. 0
Additionally, a rule is associated with will increase by 15% if the
1400 1500 1600 1700a sensor s p 2 Si ca-
1800
pable of sensing the state property p. Finally, in a PSRs the cleaned and no change oth
Time Deadline
device can also sense its own internal states. More formally, Using the predictive mo
predict the trajectory of a
Ferdinando Fioretto, William Yeoh, Enrico Pontelli. "A Multiagent System ApproachHto Scheduling Devices in Smart Homes". AAMAS, 2017.
p= {z 2 Ai | `z = `Rp ^ 9a 2 Az : p 2 z (a)} (4) steps given a schedule of a
p= {z 2 Ai | z = `Rp _`z = `R119
H
p
^9a 2 Az : p 2 z (a)} (5) Definition 3 (Predicted St
SMART HOMES   Example 2 The scheduling rule (1) describes an ASR
8
DEVICE SCHEDULING
> min(20, 100 !)
>
<
defining a goal state where the living room floor is at least 25
75% clean (i.e., at least 75% of the floor is cleaned by a > 0
vacuum cleaning robot) by 1800 hrs: >
:
?
and scheduling rules (2) and (3) describe PSRs stating that will increase by 20% if it is
S the battery
R Rcharge
C of the
C vacuum
R R C robot
cleaning R zvDevice
needs to
Schedule decrease by 25% if it is ru
be between 0 and 100 % of its full charge at all the times: charge, and no change if it
0 15 30 30 30 45 60 60 75 Cleanliness (%)
zv battery charge 0 always (2) Example 4 Consider the

65 40 15 35 55 30 5 25 0 Battery Charge (%) room, where the only actu
zv battery charge  100 always (3) a vacuum cleaning robot z
Goal 75
$ $ $ [t !t$]
For a home hi , we denote with Rp a b a scheduling rule
… 75
sume the room is 0% clean
robot zv is ⇠zt v at that time
60 over a state property p 2 PH [PZ , and time interval [ta , t60 b ].
Battery Charge (%)

Cleanliness (%)
Each scheduling rule indicates a goal state (either … a desired t+1

(0, ⇠zt v ) = 0 +
45
state of a home 8:00if it9:00 is an ASR10:00 11:00 state of a de-
or a required 45
cleanliness
⇢
min
30 vice or a home if it is a PSR) at a location `Rp 2 Li [ 30 Zi t
cleanliness (!, ⇠zv ) =
of a particular
real-time state property
energy that mustschema
p price hold over the time 0
15
interval [ta , tb ] ✓ T. Each rule is associated with a set 15of
0 actuators p ✓ Ai that can be used to reach the goal state. 0
Additionally, a rule is associated with will increase by 15% if the
1400 1500 1600 1700a sensor s p 2 Si ca-
1800
pable of sensing the state property p. Finally, in a PSRs the cleaned and no change oth
Time Deadline
device can also sense its own internal states. More formally, Using the predictive mo
predict the trajectory of a
Ferdinando Fioretto, William Yeoh, Enrico Pontelli. "A Multiagent System ApproachHto Scheduling Devices in Smart Homes". AAMAS, 2017.
p= {z 2 Ai | `z = `Rp ^ 9a 2 Az : p 2 z (a)} (4) steps given a schedule of a
p= {z 2 Ai | z = `Rp _`z = `R120
H
p
^9a 2 Az : p 2 z (a)} (5) Definition 3 (Predicted St
3 Scheduling of Devices in Smartx2 Buildings Ui =
the provider sets energyz price
j 2Zi
Through the proliferation
a1 x1 ofxsmart devices (e.g.,forsmart
i < j ther-
SMART HOMES  
2
x1 ing schema specified at regula
mostats, smart lightbulbs, smart washers, etc.)x in xour Costs
x4 i j
homes horizon
The SBDS H. We problemuse T is= the{1, .
and offices, building 0 0 20
a2 automation
x3 x4 within the larger smart grid intervals
vices of eachandbuilding ✓ : T ! in theR+
0the automated
1 8
DEVICE SCHEDULING
is becoming inevitable. Building automation is associated
fashion so as to with the pricing
minimize thescm
x3 1 0 10 theminimize
cost per kWh of energy co
control of the building’s devices with thex objective of3 im- time, the discomfort
a3 x5 x6 6 1 1
Within each smart buildin
proved comfort of the occupants, improved energy efficiency, objective optimization problem
electric devices Zi networke
and reduced operational costs. In this paper, we are interested tiveshome
intoautomation
a single objective
(a) Constraint Graph x 5
(b) Constraint Cost Table system. Alt
in the scheduling devices in smart buildings in a decentral- sum:(i.e., they cannot be stopped X X o
Domains Environmental 
ized way, where each user is responsible for the schedule of szj and minimizezj to denote the sta
Figure 1: Example
the devices in her building, under the assumption that each variables
DCOP in exact multiples oft2T timehiin2
user cooperate to ensure that the total energy consumption of zj 2↵Ziand . ↵ are weights
A solution is a value assignment to a set of variables where
the neighborhood is always within some maximum threshold The energyuconsumption oU
c
X ✓ X that Actionis consistent
State with the variables’
Pproperty domains. The such that ↵c + ↵u = 1. The S
Power (kW/h)
that is defined each hour that it is on. It wi
cost function FP ( ) = f2F,xf ✓X f( ) is the utility
by the energy provider such as a energy sum of the the following constraints:
company. costs of allrun cleanliness,
is off-the-shelf. We use the in
the applicable constraints in0.0. A solution is said
We nowto provide a description the state1 ofszthe device
T zjzjat ti
be complete = Xofcharge
if X battery isthe
theSmart
value Building
assignmentDe- for all j
vices Scheduling (SBDS)
charge problem.
battery We
charge describe
0.26 related so-8 exclusively
X when the device Cz
variables. The goal is
Example 2 The scheduling rule (1) describes an ASR to find an optimal complete solution
min(20, 100 !) if ⇠z= t t
= charge ^ ! <
> v⇢ z
lution approaches
⇤
x =where
defining a goal state argmin in Section
F (x).
the living 6. An SBDS
room floor is at least problem is com- >
< z j
tif ⇠zv = run
t 1j ^if! > szj2
stopx P 0.0 25
=
75% posed of a neighborhood
clean (i.e., atFollowing Fioretto
least 75% of H of
the floorsmart
et al.is buildings
[2016b],
cleaned by a hi 2 Hthethat
we introduce follow- t2T z
ifj ⇠zv = stop
t 0 otherw
vacuum
>
>
:
0 X
arecleaning
able toingrobot) by 1800 hrs:
definitions:
communicate with one another and whose energy ? t
otherwise
P  ` t
demands are served by an energy provider. We assume that Additionally, the executio
i
living Definition 1 For each agent ai 2A, Li ={xj 2InXother
room cleanliness 75 before 1800 (1) | ↵(xwords,
j )= atby eacha hcost
i 2H and
time step, athe discomfort
charge of v
ai } is(2)theandset(3)ofdescribe
and scheduling rules its localPSRs
variables. i = {xj 2will
stating Ithat | 9xk 2by 20%
Liincrease if it is charging
monetary expenseuntil it isuse
for the ful
the battery chargeX of thesvacuum
^ 9f 2F : ↵(x cleaning
k ) 6= arobot
i ^ {xzvj ,needs to s } isdecrease
xk }✓x f
the set ofbyits25% ifand
it is running tuntil it has less
we use Ci to denote the
be between 0 and 100 % of its
interface variables.full charge at all the times: Hard constraints
charge, and no change if it is stopped.
AAMAS-19 Tutorials at time
hiFioretto, stepTran-Thanh,
Picard, t, expressed
Rust as
Example 4 Consider the cleanliness state pro
zv battery charge
Definition 2 For each agent a 2A, its(2)
0 always
local constraint graph t
X
3 Scheduling of Devices in Smartx2 Buildings Uit = t
zj · µzj (t)
SMART HOMES  
the provider sets energy prices according
zj 2Zi
Through the proliferation
a1 x1 ofxsmart
2 devices
x1 (e.g.,forsmart
i < j ther- ing schema specified at regular intervals t
mostats, smart lightbulbs, smart washers, etc.) x4 i j
x in xour Costs
homes horizon
The SBDS H. We use T is
problem = the . . , H} to of
{1, .problem de
and offices, building 0 0 20
a2 automation
x3 x4 within the larger smart grid intervals
of eachandbuilding to represent
R neighborhoo
+
vices ✓ : T ! in the
DEVICE SCHEDULING
is becoming inevitable. Building automation x3
control of the building’s devices with thex objective
proved comfort of the a3 x5 x6 6
0
1 0
1
occupants, improved energy efficiency,
1
is the automated
1
8
10
of3 im-
associated
fashion
the
time,
objective
so as to
cost
minimizeper
with
the
the pricing
minimize
kWh of energy
discomfort
theschema
monetary
consumed
Within each smart building hi , there
optimization problem, we comb
of
adopte
users. by
co
Wa
electric devices Zi networked together an

and reduced operational costs. In this paper, we are interested tiveshomeintoautomation
a single objective
(a) Constraint Graph x 5
(b) Constraint Cost Table system. Allthrough the
the devices
in the scheduling devices in smart buildings in   a decentral-
Objective sum:(i.e., they cannot be stopped X X once they are
ized way, where each user(soft is responsible for the
constraints) Objective schedule of szj and minimize
(softzj to constraints)
denote the start time ↵c · C t
i +
and d
Figure 1: Example DCOP
the devices in her building, under the assumption that each in exact multiples oft2T timehiintervals), resp
2H
user cooperate to ensure that the total energy consumption of zj 2↵Ziand . ↵ are weights in the open
A solution is a value assignment to a set of variables where
the neighborhood is always within some maximum threshold The energyuconsumption of
c Ui each 0.3 device
0.2 3.
X ✓ X that is consistent P with the variables’ domains. The such that
each hour ↵ + ↵ = 1. The SBDS
c that uit is on. It will not consum proble
that is defined by the energy
cost function FP ( ) provider
= f2F,xsuch asf(a energy
is the utility
sum of the
f ✓X ) the isfollowing constraints:
off-the-shelf. We use the indicator funct
company. costs of all the applicable constraints in . A solution is said
We nowto provide a description
be complete if X = Xofisthe theSmart
value Building
assignmentDe-
$ the$state
for all 1 ofszthe
exclusively
j $T zjz$jat time…
device
X when the device Czj is on at tim
step
8ht,i 2an
vices Scheduling
variables.(SBDS)
The goalproblem.
is to findWe describecomplete
an optimal related solution
so- t ⇢ i
lution approaches
⇤ in Section 6. An SBDS problem is com- zj = zj 8hi 2
x = argmin F
x P (x). 1 if s  …t ^ szj +
posed of a neighborhood
Following FiorettoH of smart
et al. buildings
[2016b], hi 2 Hthethat
we introduce 8:00
follow- 9:00t2T
X j
t
z10:00
= 11:00
0 otherwise
z j
are able toing definitions: with one another and whose energy
communicate Pit the `t execution of device z
demands are served1byForaneach
energy real-time energy Additionally, price schema
Definition agentprovider.
ai 2A, LWe assume that
i ={xj 2 X | ↵(xj )= by a cost and a discomfort value. The c
hi 2H
ai } is the set of its local variables. Ii = {xj 2 Li | 9xk 2 monetary expense for the user to schedule
X ^ 9fs 2F : ↵(xk ) 6= ai ^ {xj , xk }✓xfs } is the set of its and we use Cit to denote the aggregated c
interface variables. hi at time step t, expressed as:
Definition 2 For each agent ai 2A, its local constraint graph Cit = Pit · ✓(t),
Gi = (Li , EFi ) is a subgraph of the constraint graph, where
fj where X

05 MAS Distributed Constraint Optimization

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

05 MAS Distributed Constraint Optimization

Uploaded by

Copyright:

Available Formats

Multiagent Systems

Distributed Constraint Optimization

Chapter 12: Distributed Constraint Handling and

A. Farinelli1 M. Vinyals1 A. Rogers2 N.R. Jennings2

Multi-Agent Systems. MIT Press, 2012 (2nd edition), edited by

Chapter 12: Distributed Constraint Handling and Optimization

Pervade our everyday lives

Chapter 12: Distributed Constraint Handling and Optimization

From a computational point of view, they:

Chapter 12: Distributed Constraint Handling and Optimization

Many different disciplines deal with hard computational problems that

Planning Operational Automated Reasoning Computer

Chapter 12: Distributed Constraint Handling and Optimization

Constraint Processing in Multi-Agent Systems

Focus on how constraint processing can be used to address

A set of agents must come to some agreement, typically via some

Chapter 12: Distributed Constraint Handling and Optimization

Distributed Constraint Optimization Problems (DCOPs)

Each agent negotiates locally with just a subset of other agents

Chapter 12: Distributed Constraint Handling and Optimization

A constraint network N is formally defined as a tuple hX , D , C i where:

X = {x1 , . . . , xn } is a set of discrete variables;

Chapter 12: Distributed Constraint Handling and Optimization

A hard constraint Cih is a relation Ri that enumerates all the valid

Chapter 12: Distributed Constraint Handling and Optimization

A soft constraint Cis is a function Fi that maps every possible joint

Chapter 12: Distributed Constraint Handling and Optimization

Binary Constraint Networks

Chapter 12: Distributed Constraint Handling and Optimization

Different objectives, different problems

Constraint Satisfaction Problem (CSP)

Objective: find an assignment for all the variables in the network

Constraint Optimization Problem (COP)

Objective: find an assignment for all the variables in the network

Global function = aggregation (typically sum) of local functions.

Chapter 12: Distributed Constraint Handling and Optimization

Distributed Constraint Reasoning

Chapter 12: Distributed Constraint Handling and Optimization

Distributed Constraint Reasoning

Two types of decentralized problems:

Here, we focus on DCOPs.

Chapter 12: Distributed Constraint Handling and Optimization

Distributed Constraint Optimization Problem (DCOP)

A DCOP consists of a constraint network N = hX , D , C i and a set of

Chapter 12: Distributed Constraint Handling and Optimization

Distributed Constraint Optimization Problem (DCOP)

Agents are assumed to be fully cooperative

Chapter 12: Distributed Constraint Handling and Optimization

Chapter 12: Distributed Constraint Handling and Optimization

Real World Applications

Many standard benchmark problems in computer science can be

Chapter 12: Distributed Constraint Handling and Optimization

Chapter 12: Distributed Constraint Handling and Optimization

Graph coloring - CSP

Nodes can take k colors

Chapter 12: Distributed Constraint Handling and Optimization

Graph coloring - Max-CSP

Minimize the number of conflicts

Chapter 12: Distributed Constraint Handling and Optimization

Graph coloring - COP

Different weights to violated constraints

Chapter 12: Distributed Constraint Handling and Optimization

Graph coloring - DCOP

Chapter 12: Distributed Constraint Handling and Optimization