26th Nov 2001Univ. Nebraska1 Advanced Scheduling and Optimization: Cutting the Costs of...
-
date post
19-Dec-2015 -
Category
Documents
-
view
213 -
download
0
Transcript of 26th Nov 2001Univ. Nebraska1 Advanced Scheduling and Optimization: Cutting the Costs of...
26th Nov 2001 Univ. Nebraska 1
Advanced Scheduling and Optimization: Cutting the Costs of Manufacturing
Brian Drabble
Computational Intelligence Research Laboratory
www.cirl.uoregon.edu
&
On Time Systems, Inc
www.otsys.com
26th Nov 2001 Univ. Nebraska 2
Overview
• Constraint based scheduling• Algorithms
– LDS and Schedule Pack– Squeaky Wheel Optimization
• Applications– Aircraft assembly– Ship construction
• Future Directions• Summary
26th Nov 2001 Univ. Nebraska 3
Constraint Based Scheduling
• Problem characteristics• Search based techniques
26th Nov 2001 Univ. Nebraska 43
Problem Characteristics
–Task details:• resource requirements
• deadlines/release times
• value
26th Nov 2001 Univ. Nebraska 54
Problem Characteristics
–Task details–Resource characteristics:
• type• capacity• availability• speed, etc.
26th Nov 2001 Univ. Nebraska 65
Problem Characteristics
¨ Task details¨ Resource characteristics¨ Precedences:
– necessary orderings between tasks
26th Nov 2001 Univ. Nebraska 76
Problem Characteristics
–Constraints:• setup costs• exclusions• reserve capacity• union rules/business rules
¨ Task details¨ Resource characteristics¨ Precedences
26th Nov 2001 Univ. Nebraska 87
Problem Characteristics
–Constraints
–Optimization criteria:• makespan, lateness, cost,
throughput
¨ Task details
¨ Resource characteristics¨ Precedences
26th Nov 2001 Univ. Nebraska 98
Optimization Techniques
• Operations Research (OR)– LP/IP solvers
• seem to be near the limits of their potential
• Artificial Intelligence (AI)– search-based solvers
• performance increasing dramatically• surpassing OR techniques for many problems
26th Nov 2001 Univ. Nebraska 109
Search-based Techniques
• Systematic– explore all possibilities
• Depth-First Search• Limited Discrepancy Search
• Nonsystematic– explore only “promising” possibilities
• WalkSAT• Schedule Packing
26th Nov 2001 Univ. Nebraska 1110
Heuristic Search
– A heuristic prefers some choices over others– Search explores heuristically preferred options
26th Nov 2001 Univ. Nebraska 1211
Limited Discrepancy Search
– Better model of how heuristic search fails
26th Nov 2001 Univ. Nebraska 1312
Limited Discrepancy Search– LDS-n deviates from heuristic exactly n times
on path from root to leaf
LDS-1LDS-0
26th Nov 2001 Univ. Nebraska 1413
Schedule Packing– Post-processing to exploit opportunities
1 1
2 2
26th Nov 2001 Univ. Nebraska 1514
Schedule Packing– schedule longest chains first
• starting from right
1 1
2 2
1 1
2 2
26th Nov 2001 Univ. Nebraska 17
Squeaky Wheel Optimization
Construct
Mission 1234
AAR 234
SEAD 34
Mission 4567
26th Nov 2001 Univ. Nebraska 18
Squeaky Wheel Optimization
Analyze
“High attrition rate”
“Outside target time window”
“Low success rate”
“Not attacked”
Scalability
0
5
10
15
20
25
0 50 100 150 200 250 300Number of Tasks
% Over Best Solution
TABULP/IPSWO
26th Nov 2001 Univ. Nebraska 2417
Aircraft Assembly
McDonnell Douglas / Boeing
– ~570 tasks, 17 resources, various capacities
– MD’s scheduler took 2 days to schedule
– needed:
• better schedules (1 day worth $200K–$1M)
• rescheduler that can get inside production cycles
26th Nov 2001 Univ. Nebraska 2518
Problem Specification
– Task/precedence specification• mostly already existed for regulatory reasons
26th Nov 2001 Univ. Nebraska 2619
Problem Specification
– Task/precedence specification• mostly already existed for regulatory reasons
– Resource capacity profiles• labor profile available from staffing information• others determined from SOPs, etc.
26th Nov 2001 Univ. Nebraska 2720
Problem Specification
– Task/precedence specification• mostly already existed for regulatory reasons
– Resource capacity profiles• labor profile available from staffing information• others determined from SOPs, etc.
– Optimization criterion• simple makespan minimization
26th Nov 2001 Univ. Nebraska 2821
Problem Specification
– Task/precedence specification• mostly already existed for regulatory reasons
– Resource capacity profiles• labor profile available from staffing information• others determined from SOPs, etc.
– Optimization criterion• simple makespan minimization
– Solution checker• available from in-house scheduling efforts
26th Nov 2001 Univ. Nebraska 2922
The Optimizer
• LDS to generate seed schedules• Schedule packing to optimize
– intensification improves convergence speed
• etc.
26th Nov 2001 Univ. Nebraska 3023
Performance
– ~570 tasks, 17 resources, various capacities• about 1 second to first solution• about 1 minute to within 2% of best known• about 30 minutes to best schedule known
26th Nov 2001 Univ. Nebraska 3124
Performance
– ~570 tasks, 17 resources, various capacities• about 1 second to first solution• about 1 minute to within 2% of best known• about 30 minutes to best schedule known
– 10-15% shorter makespan than best in-house• 4 to 6 days shorter schedules
26th Nov 2001 Univ. Nebraska 3225
Performance
– ~570 tasks, 17 resources, various capacities• about 1 second to first solution• about 1 minute to within 2% of best known• about 30 minutes to best schedule known
– 10-15% shorter makespan than best in-house• 4 to 6 days shorter schedules
– 2 orders of magnitude faster scheduling• scheduler runs inside production cycle• less need for rescheduler
26th Nov 2001 Univ. Nebraska 3326
Extensions
Boeing:– multi-unit assembly– interruptible tasks– persistent assignments– multiple objectives
• e.g., time to first completion, average makespan, time to completion
• fast enough to use for “what-iffing”– discovered improved PM schedule
26th Nov 2001 Univ. Nebraska 3427
Submarine Construction
General Dynamics / Electric Boat– 7000 activities per hull, approx 125 resources– Electric Boat’s scheduler takes 6 weeks– needed:
• cheaper schedules• faster schedules of contingencies
26th Nov 2001 Univ. Nebraska 35
Problem Specification
• reschedule shipyard operations to reduce wasted labor expenses
• efficient management of labor profiles– reduce overtime and idle time– hiring and RIF costs
26th Nov 2001 Univ. Nebraska 36
Optimizer
• ARGOS is new technology developed specifically with these goals in mind
26th Nov 2001 Univ. Nebraska 37
• Labor costs of existing schedule: $155m• Time to produce existing schedule: ~6 weeks
• 15% reduction in cost, 50x reduction in schedule development time
Performance: One Boat
Iteration Time Savings
1 2 min 8.4% $13.0M7 10 min 11.4% $17.7M20 34 min 11.8% $18.2M
Ultimate ~24hrs 15.5% $24.0M
26th Nov 2001 Univ. Nebraska 38
Performance: Whole Yard
• All hulls, about 5 years of production• Estimated cost of existing schedule: $630M
• No existing software package can deal with the yard coherently
Iteration Time Savings
1 24 min 7.8% $49M7 60 min 10.2% $65M20 4 hours 10.7% $68M
Ultimate 4 days 11.5% 73M
26th Nov 2001 Univ. Nebraska 39
Extensions
• Shared resources– dry dock– cranes
• Sub-assemblies– provided by different yards and suppliers
• Repair– dealing with new jobs
26th Nov 2001 Univ. Nebraska 40
Future Applications
• Workflow management– STRATCOM checklist manager– IBM
• E-Business– supply chain management
• Military– air expeditionary forces– logistics
26th Nov 2001 Univ. Nebraska 41
Future Work
• Robustness• Distributed scheduling• Common task description
26th Nov 2001 Univ. Nebraska 42
Penalty Box Scheduling
• Sub-set of the tasks with higher probability of success.– 90% probability of destroying 90% of the targets?– 96% probability of destroying 75% of the targets?
• Inability to resource leads to a task “squeak” • Blame score related to user priority and
“uniqueness”• Reduce the target percentage until no
significant improvement is found
26th Nov 2001 Univ. Nebraska 43
Semi-Flexible Constraints
• The time constraints provided by the users tended to be ad-hoc and imprecise– heuristics based on sortie rate, no of targets, etc– this is what we did last time so it must be right!!
• Not a preference– this is what I want until you can prove otherwise!!
• Two algorithms were investigated– pointer based– ripple based
26th Nov 2001 Univ. Nebraska 44
Semi-Flexible Constraints: Pointer Based
“Attack the IAD before power system”
IAD-E
0 3000 6000
Time (Minutes)
Power-EIAD-L Power-L
26th Nov 2001 Univ. Nebraska 45
Semi-Flexible Constraints: Pointer Based
“Attack the IAD before power system”
IAD-E
0 3000 6000
Time (Minutes)
Power-EIAD-L Power-L
26th Nov 2001 Univ. Nebraska 46
Semi-Flexible Constraints: Pointer Based
“Attack the IAD before power system”
IAD-E
0 3000 6000
Time (Minutes)
Power-EIAD-L Power-L
26th Nov 2001 Univ. Nebraska 47
Semi-Flexible Constraints: Ripple Based
IAD-E
0 3000 6000
Time (Minutes)
Power-EIAD-L Power-L
“Attack the IAD before power system”
26th Nov 2001 Univ. Nebraska 48
Semi-Flexible Constraints: Ripple Based
IAD-E
0 3000 6000
Time (Minutes)
Power-E Power-L
“Attack the IAD before power system”
IAD-L
26th Nov 2001 Univ. Nebraska 49
Semi-Flexible Constraints: Ripple Based
IAD-E
0 3000 6000
Time (Minutes)
Power-E Power-L
“Attack the IAD before power system”
IAD-L
26th Nov 2001 Univ. Nebraska 50
Semi-Flexible Constraints: Ripple Based
IAD-E
0 3000 6000
Time (Minutes)
Power-LPower-E
“Attack the IAD before power system”
IAD-L
26th Nov 2001 Univ. Nebraska 51
Common Task Model
Plan Ready Fly Execute Recover “Drop 120, MK-84s from 3 B-52s at location X,Y at 22.00 on D+5”
30 mins 20 mins 40 mins 5 mins 60mins
P R F E RB-52 Flight
P R F E RAAR
P R F E RAWACS
P R F E RBomb Depot
P R F E RCAP Flight Information & Control
P R F E RWeapon Loader
P R F E RSEAD Flight
26th Nov 2001 Univ. Nebraska 52
Example Problem (2)
• The AWACS aborts on take off!
P R F E RB-52 Flight
P R F E RAAR
P R F E RAWACS
P R F E RBomb Depot
P R F E RCAP Flight
P R F E RWeapon Loader
P R F E RSEAD Flight
26th Nov 2001 Univ. Nebraska 53
Summary
Advances in search technology: Tasks Resources Type Feasible?
– 1993: 64 6 Job Shop X – 1996: ~570 17 RCPS barely– 1999: 1000s dozens RCPS – 2001: 10000s hundreds RCPS
• Search works!– search-based technology has matured– large, real-world, problems are solvable– tech-transfer path is short