Considering Feedback Loops in Constraint Programming … · 2018-09-14 · Based on recent projects...
Transcript of Considering Feedback Loops in Constraint Programming … · 2018-09-14 · Based on recent projects...
Considering Feedback Loops inConstraint ProgrammingMethodology
Helmut Simonis
ModRef 18, August 27, 2018
Based on work with...
• Current and past staff at Insight
• ICON project partners
• TASC team in Nantes
• Constraint team in Uppsala
Insight Centre for Data Analytics Slide 2ModRef 18, August 27, 2018
A Slightly More Accurate Picture
Insight Centre for Data Analytics Slide 4ModRef 18, August 27, 2018
Key Points
• This is not a realistic model
• Definitely not for industrial work
• Nobody knows what the problem is
• The problem keeps changing
• The process is a set of nested feedback loops, allinfluencing each other
Insight Centre for Data Analytics Slide 5ModRef 18, August 27, 2018
Focus on Benchmarking Hides This Issue
• Work on clearly defined problem
• Solutions may already exist for comparison• Quality is easily defined
• Better objective value• Faster than previous approaches
• Only run-time counts• Should be:
• Overall development time until first usable solution• Including design, revision, training
Insight Centre for Data Analytics Slide 6ModRef 18, August 27, 2018
Some Real-World Examples
• Based on recent projects at Insight
• Not in production use
Insight Centre for Data Analytics Slide 7ModRef 18, August 27, 2018
Example: Datacenter Management
Electricity PriceWeather Forecast
VirtualMachines
DemandForecast
CurrentAssignment
VMAllocator Resources
VM Allocation
ManagementPlatform
Running VM
Insight Centre for Data Analytics Slide 8ModRef 18, August 27, 2018
Example: Transport
ExpectedTravel Times
TransportRequests
CurrentLocation
TransportSolver
Work RulesPreferences
Planned Tours
FleetOperation
Actual Tours
Customers
Manager
Agents
Insight Centre for Data Analytics Slide 9ModRef 18, August 27, 2018
Example: Personnel Rostering
UnplannedUnavailabilityEmergencies
DemandPlan
History RosteringSolver
Work RulesPreferences
Planned Schedule
Operations
Actual Schedule
Manager
Staff
Insight Centre for Data Analytics Slide 10ModRef 18, August 27, 2018
The General Scheme
ExogenuousData
DecisionVariables
CurrentState
ConstraintModel
ConstraintsPreferencesObjectives
Plan
Realization
Actuals
Insight Centre for Data Analytics Slide 11ModRef 18, August 27, 2018
A Blueprint for Interaction
• Developed in the European ICON project
• Partners KU Leuven, Montpellier, Pisa, UCC
• Ways of combining Machine Learning with CP
Insight Centre for Data Analytics Slide 13ModRef 18, August 27, 2018
DecisionCamp 2017, Eric Mazeran, IBM
Insight Centre for Data Analytics Slide 14ModRef 18, August 27, 2018
Example: Intra-City Transport
Trafficin City
TrafficAnalysis
TrafficMatrix
Public TransportRoute
Optimization
ChangedServices
• Optimization only as good as data feeding into it
Insight Centre for Data Analytics Slide 15ModRef 18, August 27, 2018
Feedback May Lead to More Traffic
Trafficin City
TrafficAnalysis
TrafficMatrix
InvestmentOptimization
Buildroad
• The ‘‘world’’ reacts to changes
• That may be difficult to predict
Insight Centre for Data Analytics Slide 16ModRef 18, August 27, 2018
Reacting to Real-time Electricity Prices
ElectricityMarket
PriceForecasting
Real-timePrice
Forecast
EnergyCost AwareScheduling
• Good way to optimize cost individually
• May lead to oscillation if everybody does it
Insight Centre for Data Analytics Slide 17ModRef 18, August 27, 2018
Feedback Loops in Modelling
Problem
User
(Informal)Description
Modeler
Model
Solver
ProposedSolution
Designer
IntegrationConstraint
Bias
Insight Centre for Data Analytics Slide 18ModRef 18, August 27, 2018
The Future: Automated Modelling
Problem
Sample/HistoricalSolutions
ModelSeeker
ModelElements
User/Modeler
Model
Solver
ProposedSolution
SolverGenerator
IntegrationConstraint
Bias
Insight Centre for Data Analytics Slide 19ModRef 18, August 27, 2018
ModelSeeker
• Learn elements of constraint models from positiveexamples
• Highly stuctured problems• Essentially matrix models• Conjunctions of global constraints• Based on global constraint catalog
• System suggests sets of constraints
• Relies on user to select meaningful subset
Insight Centre for Data Analytics Slide 20ModRef 18, August 27, 2018
SolverGenerator
• Generate constraints for families of constraintsautomatically
• Currently being developed for time-series
• No limit on number of different constraints needed tomodel problem
• Also generate implied constraints for conjunction of basicconstraints
Insight Centre for Data Analytics Slide 21ModRef 18, August 27, 2018