objectives and approaches
DESCRIPTION
Flexible Methods for Multi-agent Distributed Resource Allocation by Exploiting Phase Transitions Modeling and Analyzing Resource Allocation Problems Using Soft Constraint Satisfaction and Optimization. - PowerPoint PPT PresentationTRANSCRIPT
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 1
Flexible Methods for Multi-agent Distributed Resource Allocation by Exploiting Phase Transitions
Modeling and Analyzing Resource Allocation Problems Using Soft Constraint Satisfaction and Optimization
Weixiong Zhang (PI) Kenneth Swanson, Xiaotao Zhang,
Peng Wang, Michael P. Moran,Guandong Wang, Zhao Xing,
Zhongshen GuoComputational Intelligence Center and
Computer Science DepartmentWashington University in St. Louis
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 2
Objectives and Approaches
• Understanding and characterizing resource allocation problems in ANTs applications.– Modeling methods: soft constraint satisfaction/optimization– Phase transitions and backbones (sources of complexity)– Scalability (e.g. impact of problem structures)
• Developing general and efficient algorithms for resource allocations– Systematic search methods– Approximation methods– Distributed algorithms– Phase-aware problem solving for good enough/sooner
enough solutions
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 3
Work in this period
• EW challenge problem– Design and develop a moving target tracking
system in RadSim
– Preliminary working system
– Testbed for studying many technical difficulties
– (more to come at next PI meeting)
• Marbles pilot scheduling problems– (main focus of this presentation)
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 4
Current Work on EW Challenge Problem
• Technical issues under consideration– Scalability
• how problem structures and agent organization affect complexity
– Uncertainty in resource conflict resolution: uncertainty in measurement, communication error, etc.
– Scan scheduling for detecting new targets quickly with small amount of energy
– Irregular sensor layout• We have shown that triangle topology provides the best area
coverage• What if sensor layout is out of your control – how to quickly
form teams
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 5
Marbles Scheduling Problem
• Main focus of this period
• Some results
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 6
The Marbles Problem• Resource allocation in a task scheduling problem
– Schedule as many tasks as possible (to reduce the overall penalty of unscheduled tasks)
– Block resource requirement• Each task requires a set of resources• It cannot be schedule unless all resource requirements are fulfilled
– Exclusive resource contention• A shared resource may be applicable to multiple requirements• But it can be used to fulfill only one requirement
R1 R2 R3
T1
Q1,1 0 1 1
Q1,2 1 0 0
T2
Q2,1 1 1 0
Q2,2 1 1 0
Resource requirements
Resources
Tasks
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 7
The Problem is Difficult
• The problem is NP-hard– The decision version is NP-complete– Reduced from set packing (NP-complete)
• Set packing: – Given a collection S of finite sets of elements, a positive
integer K
– Decide: if S contains at least K mutually disjoint subsets
• Reduction:– Map an elements to a resource
– Map a subset to a task
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 8
Technical Content
• Hard and soft constraints• Modeling consideration and choices• Constraint models
– Models in optimization
– Models in satisfaction
• Experimental analysis (phase transitions)• Current and future work
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 9
Hard and Soft Constraints
• Task constraints (soft constraints) – Ctask: turning on tasks (typically, not all of them can be
satisfied at once)• Constraint (Ti = 1) to represent turning on task Ti
• Weight equal to 1 or its penalty
• Block resource requirements (hard constraints) – Creq: Fulfilling resource requirement of a task if it is on– Weight is more than the total weight of soft constraints
• Exclusive resource contention (hard constraints) – Cres: A resource can only be used by one requirement– Weight is more than the total weight of soft constraints
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 10
Main Modeling Considerations• Optimization vs. decision
– Optimization: try to turn on all tasks, and then find the maximal number of tasks that can be indeed turned on
– Decision: Guess the possible number of tasks that can be turned on, and then verify it. Do a binary search on the number of tasks. (caution: it may not work if tasks are weighted and it is to minimize the overall weight of unscheduled tasks.)
• General variables versus Boolean variables– CSP/COP (Constraint Optimization Problem) versus SAT/MAX-SAT– K-encoding issue
• Which choice to take and under what conditions?
Optimization Decision
General variables A COP model A set of CSP models
Boolean variables A MAX-SAT model A set of SAT models
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 11
Main Modeling Choices
• Variable versus values– Resources as variables and requirements as values– Or vice versa– Which one to use?
R1 R2 R3
T1
Q1,1 0 1 1
Q1,2 1 0 0
T2
Q2,1 1 1 0
Q2,2 1 1 0
Resource requirements
Resources
Tasks
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 12
Main Modeling Choices
• Expressiveness of a model• E.g. Two resources may be assigned to one
requirement (but one is used)• Should hidden constraints be made explicit?
• Interaction between modeling considerations and choices and search algorithms
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 13
COP/CSP Models
Optimization Decision
COP1(requirements as variables)
CSP1(requirements as variables)
COP2(resources as variables)
CSP2(resources as variables)
COP3(resources as variables,
more explicit than COP2)
CSP3(resources as variables,
more explicit than CSP2)MAX-SAT4 SAT4
MAX-SAT5(more explicit than MAX SAT4)
SAT5(more explicit than SAT4)
Optimization Decision
General variables A COP model A set of CSP models
Boolean variables A MAX-SAT model A set of SAT models
Original ISI Marbles model
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 14
COP2 Model: Resources as Variables
t: # of tasksqi: # of resource requirements of task ir: # of resourcesTi: Boolean variable for task iRk = { Qij | task i, requirement j }
Ctask = ^k=1..t (Ti = 1)
Cblock = ^k=1..t Cblock(Ti)
Cblock(Ti)= ^j=1..qi ((Ti=0) Vk=1..r(Rk=Qij))
R1 R2 R3
T1
Q1,1 0 1 1
Q1,2 1 0 0
T2
Q2,1 1 1 0
Q2,2 1 1 0
Ctask = (T1 = 1) ^ (T2 = 1) (T1=0) V(R2=Q11) V(R3=Q11)Cblock = (T1=0) V(R1=Q12) (T2=0) V(R1=Q21) V(R2=Q11) (T2=0) V(R1=Q22) V(R2=Q22)
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 15
COP3 Model: More Explicit than COP2t: # of tasksqi: # of resource requirements of task ir: # of resourcesTi: Boolean variable for task iRk = { Qij | task i, requirement j}
Ctask = ^k=1..t (Ti = 1)
Cblock = ^k=1..t Cblock(Ti) ^k=1..t C'block(Ti)
Cblock(Ti)= ^j=1..qi ((Ti=0) Vk=1..r(Rk=Qij))
C'block(Ti) = ^((Ru ≠ Qij) V (Rv ≠ Qij))
R1 R2 R3
T1
Q1,1 0 1 1
Q1,2 1 0 0
T2
Q2,1 1 1 0
Q2,2 1 1 0
Ctask = (T1 = 1) ^ (T2 = 1) (T1=0) V(R2=Q11) V(R3=Q11) Cblock = (T1=0) V(R1=Q12) (T2=0) V(R1=Q21) V(R2=Q11) (T2=0) V(R1=Q22) V(R2=Q22) (R2 ≠ Q11) V(R3 ≠ Q11)C'block = (R1 ≠ Q21) V(R2 ≠ Q21) (R1 ≠ Q22) V(R3 ≠ Q22)
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 16
Phase Transitions
• Marbles problems (8 tasks with 2 requirements each)
02
46
86
7
8
9
100
0.2
0.4
0.6
0.8
1
# tasks to be turned on
8 tasks, 2 resource requirement/task, density 0.3
# resources
poss
ibili
ty t
hat
task
s ca
n be
tur
ned
on
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 17
Phase Transitions (2)
• Marbles problems (8 tasks with 2 requirements each)
0.10.15
0.20.25
0.30.35
6
7
8
9
100
0.2
0.4
0.6
0.8
1
density
8 tasks, 2 resource requirement/task
# resources
poss
ibili
ty 4
tas
ks t
urne
d on
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 18
Experimental Results:Systematic search and local search
Optimization Decision
COP1(requirements as variables)
CSP1(requirements as variables)
COP2(resources as variables)
CSP2(resources as variables)
COP3(resources as variables,
more explicit than COP2)
CSP3(resources as variables,
more explicit than CSP2)MAX-SAT4 SAT4
MAX-SAT5(more explicit than MAX SAT4)
SAT5(more explicit than SAT4)
Optimization Decision
General variables A COP model A set of CSP models
Boolean variables A MAX-SAT model A set of SAT models
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 19
Experimental Results: Complete Algorithm
• MAX SAT Models
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 20
Experimental Results: Complete Algorithm
• SAT Models
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 21
Experimental Analysis: Complete Algorithm
• Summary
Optimization Decision
Model 4 3
MAX SAT4
2
SAT4
Model 5 4
MAX SAT5
1
SAT5
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 22
Experimental Analysis: Local Search
• Experiment setup– A WalkSAT-like algorithm for all models– Given all models the same total amount of CPU time
(adjusted by the numbers of moves and restarts– Measure the final solution quality– CPU time the best solution found the first time
• Results– Optimization models are better than decision models– COP1 is the best– COP2 is better than COP3
Optimization Decision
COP1(requirements as variables)
CSP1(requirements as variables)
COP2(resources as variables)
CSP2(resources as variables)
COP3(resources as variables,
more explicit than COP2)
CSP3(resources as variables,
more explicit than CSP2)
MAX-SAT4 SAT4
MAX-SAT5(more explicit than MAX SAT4)
SAT5(more explicit than SAT4)
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 23
Summary
• Marbles problems are indeed difficult
• There are phase transition phenomena in Marbles problems
• Modeling and search algorithms affect each other
• Good modeling methods can greatly reduce problem-solving time
ANTs PI meeting, Dec. 17, 2001Washington University / DCMP 24
Next Steps
• Marbles scheduling problem– More accurate results on phase transitions– More efficient search algorithms– Large problems– Timed Marbles problems
• for a long period, e.g., days,weeks and months
• EW challenge problem– Scalability
– Uncertainty
– Scan scheduling (larger coverage, less energy)
– Irregular layout