Upload
kenneth-willis
View
213
Download
0
Embed Size (px)
Citation preview
P MAll rights reserved (Tsang et al)
BT’s Collaborator ProblemRegions
Engineers
Jobs
P
Controllers
Tim
e
Job
Dur
atio
n /
Eng
inee
r A
vaila
bilit
yManager
Research Agenda:
To define for management a mechanism to achieve all-win solutions
P MAll rights reserved (Tsang et al)
Flexible Workforce Scheduling
for
Staff-empowerment
Edward Tsang, Qingfu Zhang, Wudong Liuand Tim Gosling
Botond Virginas, Chris Voudouris and G Owusu
All rights reserved; no part of this presentation should be copied or distributed without prior consent by the authors
P MAll rights reserved (Tsang et al)
The RECONNET Model
Manager
SellersBuyers
Solve
Req
uest
,
Contro
l Coe
ffici
ents Solution R
eports
Invitation to bids
Bids
Contract RetractionOffers / Acceptance / Reject / Wait
(Example)
M
Impl
emen
tatio
nR
esul
ts
Offers / rejects (binding)
Accept / Reject Offers (binding)
Multi-objectiveoptimisation
P MAll rights reserved (Tsang et al)
RECONNET Buyers
Region
Buyer = Controller ElementFor Completing Jobs
Job = {Start Time, Duration,Skill Required, Position,Value}
Reports
Control Information
Revenue, Distance,Preference & IncompletionCoefficients
Job Complete /Incomplete,Distance, Preference
P MAll rights reserved (Tsang et al)
RECONNET Sellers
Region
Seller = Controller Element ForAssigning Engineers
Engineer = {Position, Availability, Skill Preference}
Reports
Control Information
Assignment, Distance,Redundancy, Load Balancing
Contracts:Job,Distance, Preference, EngineerEngineers:Usage
Availability = Times Engineer AvailableSkill Preference = Skills in preference order
P MAll rights reserved (Tsang et al)
Manager: Multi-objective Optimisation
Manager
SellersBuyersInteract to
form contracts
Repeat
Control CoefficientsBuyer: Distance, Preference,
Revenue & IncompletionSeller: Assignments, Distance,
Redundancy, Load Balancing
Solution ReportsJobs: Completion, Distance & PreferenceContracts: Distance, Preference, EngineerEngineers: Total and available time
Look after company interests
P MAll rights reserved (Tsang et al)
Building Pareto Set of solutionsW
eig
ht
sf- va
lues
Ru
n
1
W 1
,1
w 1,
2
w 1,
3f 1
,1
f 1
,2
f1
f2R
un
2
W 2
,1
w 2,
2
w 2,
3f 2
,1
f 2
,2
Ru
n
3W 3
,1
w 3,
2
w 3,
3f 3
,1
f 3
,2
Ru
n
4
W 4
,1
w 4,
2
w 4,
3f 4
,1
f 4
,2
Ru
n
5
W 5
,1
w 5,
2
w 5,
3f 5
,1
f 5
,2
• Given quality metric for Pareto sets• Attempt to learn mapping from variable space to fitness space
P MAll rights reserved (Tsang et al)
Communication System Details• Problem solved over multiple rounds
• One round consists of four phases– Phase 1: Buyer to Buyer (Retraction Requests)– Phase 2: Buyer to Seller (Option requests, offers and
rejections)– Phase 3: Seller to Buyer (Options, Offer Accept, Option
Retract/Reject)– Phase 4: Buyer to Buyer (Retraction Accept, Reject &
Wait)
• System synchronises on the end of each round to keep everyone in step & prevent unfair advantage
Comms System 2
P MAll rights reserved (Tsang et al)
Communication System
Dete
rmin
esC
ontrol P
aram
eters
Manager
Buyer
Seller
Control
Para
meters
Start
Solving
Retract
Requests
Option
R
equests,O
ption O
ffers,O
ption R
ejects Option
s (includ
ingR
etraction
based options), O
ptionA
ccepts, Option
Retracts
(rejects)
Retract
Accepts,
Retract
Rejects,
Retract W
aits
End
D
etected
Report
Request
Report
sR
eports
Phase 1 Phase 2 Phase 3 Phase 4
Round 1 Round nSolution
StartSolution
End
Four C
omm
unication Phase in E
ach Round A
s Rou
nd 1
Retraction Example
P MAll rights reserved (Tsang et al)
Implementation• Use of Weblogic for communication
– JMS
• Use of Java for solution implementation
• Distributed implementation
• File based configuration and HTML/spreadsheet friendly output
• Flexible architecture allows easy expansion and modification
P MAll rights reserved (Tsang et al)
Contract Retraction Mechanism• Allows Hill-Climbing to find better solutions
– Without it solutions are ‘one shot’
Buyer 1 Buyer 2
Seller
Possible Chain of contract release ….
Seller
(2) Bids that requires Contract-release ($20)
(3) Binding request for contract release, with a compensation offer ($60)
(4) Accept contract-release
(4) Release contract
(1) Request for bids
Unserved job ($100)
P MAll rights reserved (Tsang et al)
General Results• The system is able to generate solutions to
‘real’ sized problems in about 5 minutes
• The system’s hill climbing mechanism allows initial solutions to be improved upon
• Managers are able to effect the solutions discovered by Buyer and Sellers
P MAll rights reserved (Tsang et al)
Management Power
• Force reduction in distance at Buyers
• Force reduction in preference at Buyers
• Force reduction in distance at Sellers
• Different Buyer configurations
P MAll rights reserved (Tsang et al)
Hill Climbing Demonstrated
• Use of retraction based options for improvement
Demonstration 3
UnfulfilledFree
P MAll rights reserved (Tsang et al)
Related Work
P MAll rights reserved (Tsang et al)
Related Work• Contract Net Protocol – no backtracking
– Smith 1980, IEEE Trans Computers
• Decommitting among cooperative Agents– Sen & Durfee 1994, 96, 98
• Contingency Contract – Based on future events, Raiffa, 1982
• Leveled Commitment Protocol – Sandholm & Lesser 2001, 02
• Two-phased negotiation– Price change between phases only, Akinee 2004
P MAll rights reserved (Tsang et al)
Leveled Commitment Protocol• Self-interested agents
• Decommission by penalties– Computed before contracts become binding
• Difficult to determine in a dynamic situation
• Could find global optimal– Through systematically search
• Intractable for BT’s problem with up to 105 jobs
• Nash equilibrium can be derived– Assuming perfect information
• Not true in many real life problems
P MAll rights reserved (Tsang et al)
Conclusions• Scheduling+Management problem
– Motivated by BT’s application– Multi-objectives optimisation
• Invented protocol for distributed hill climbing– Incorporating release & pricing mechanism– General protocol, could be used to implement different
HC strategies, such as GLS, Tabu Search
• This is a practical solution:– Management control– Local ownership– Found solutions to real sized problem in 5 mins
C
P MAll rights reserved (Tsang et al)
Questions
P MAll rights reserved (Tsang et al)
Future Expansion• Buyer: Contract Ordering and Selection
– When is a retraction option the better risk?– Conflict detection and offer delay
• Seller: Contract Generation and Pruning– When to issue retraction options or not– Is there a place for multiple retraction based options?– Should all contracts be priced?
• Seller: Contraction Offer acceptance– Should we consider delays to acceptance?– How do we bundle acceptances?
• Managers– How best to control solution formation in the Pareto set– How to present options to users
P MAll rights reserved (Tsang et al)
Distributed v CentralisedCentralised
• Advantages– Strong central control– Standardisation– No comms cost in
solution finding
• Disadvantages– Little local ownership– High comms cost in
dynamic environments
Distributed• Advantages
– Local ownership– Reduced scheduling
time– Divide and conquer– Natural to problem
• Disadvantages – Difficult to manage– Difficult to enforce
standards– Comms overhead
P MAll rights reserved (Tsang et al)
Mechanism Design
• Nobel Prize in Economics 2007
• For having laid the foundations of mechanism design theory
Leonid Hurwicz
Minnesotab.1917
Eric S Maskin
Princetonb.1950
Roger B MyersonChicagob.1951