Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
www.localsolver.com
1/18
2 40
Welcome
This is the first coming of the LocalSolver team in Japan
We are so pleased to come here in TokyoWarm welcome to all of you for joining this meeting
Special thanks to our Japanese partner MSI for the organization
Frédéric GardiCo-Founder & Managing Partner
Tiphaine RougerieOptimization Engineer
3 40
Agenda
LocalSolver companyQuick tour of LocalSolverBenchmarksBusiness casesLicensing and pricingMore business cases
4 40
LocalSolverCompany
5 40
LocalSolver
Optimization Solver & Services
• Supply chain optimization
• Vehicle routing
• Production scheduling
• Workforce planning
• Network optimization
• Revenue management
• Asset management
• etc.
SOFTWARE EDITOR
SERVICE PROVIDER
15 experts in mathematics, computer science, information technology€2M in revenues, +50% grow in 2018, 5x grow since incorporation in 2012
6 40
Passionate people
Thierry BenoistCo-Founder & Managing Partner, in charge of Services
• Ecole Polytechnique (X95), PhD and Habilitation in Computer Science
• 20 years of experience in Operations Research
• Sole author of a software optimizing an investment of €1 billion (Bouygues Telecom)
• Prize ASTI of the best applied PhD thesis (2005), 3rd Prize Robert Faure ROADEF (2006)
• Finalist of the EURO Excellence in Practice Award (2012)
Frédéric GardiCo-Founder & Managing Partner, in charge of LocalSolver
• PhD and Habilitation in Computer Science (UPMC, Paris 6)
• 20 years of experience in Operations Research
• Sole author of a software used by 10,000 users in the retail banking industry (Société Générale)
• 1st Prize Junior & Senior Challenge ROADEF 2005, 2nd Prix Senior Challenge ROADEF 2007
• 1st Prize Industrial Applications ROADEF (2011), 1st Prize Robert Faure ROADEF (2012)
• President of ROADEF, the French Operations Research Society (2014-2015)
7 40
100 clients in 20 countries
100 clients50% with revenues > €1bn
from 20 countriesAustralia, Austria, Belgium, Brazil,
Canada, China, Denmark,Finland, France, Germany, Italy, India, Japan, Norway, Portugal,
Slovakia, Spain, Sweden, United Kingdom, USA
2,000 academic usersfrom 90 countries
8 40
LocalSolver
9 40
Swiss Army Knife for math optimization
All-Terrain & All-In-One
Discrete, Numerical, Black-Box
Fast & Scalable
Innovative Resolution Technology
10 40
Differentiators
New-generation optimization technology• All-in-one: combines different optimization techniques
• Innovative: integrates unique heuristic search techniques
Easier to use than any other solver• Model & Run optimization solver
• Natural mathematical modeling formalism
• One-click resolution: no need of complex tuning
• LSP: innovative scripting & modeling language for fast prototyping
• Lightweight object-oriented Python, C++, Java, C# APIs for tight integration
• Available for all platforms: Windows, Linux, macOS, Solaris, 32-bit, 64-bit
11 40
LocalSolver 9.0 coming soon
Major features• Better and faster solutions for list-based models, especially routing and scheduling problems
• Better and faster solutions for nonlinear continuous problems (NLP)
• Better and faster solutions for multiobjective models
• Better and faster lower bounds
Problems
LocalSolver 9.0
LP & MIP solvers
NLP solvers
DiscreteNumerical
CP & SAT solvers
12 40
LocalSolverQuick tour
13 40
Nonlinear optimization
Maximize the volume of a bucket with a given surface of metal
𝑟
𝑅
ℎ
𝑉 =𝜋ℎ
3(𝑅2 + 𝑅𝑟 + 𝑟2)
S = 𝜋𝑟2 + 𝜋(𝑅 + 𝑟) 𝑅 − 𝑟 2 + ℎ2
function model() {
R <- float(0,1);r <- float(0,1);h <- float(0,1);
V <- PI * h / 3.0 * (R*R + R*r + r*r);S <- PI * r * r + PI*(R+r) * sqrt(pow(R-r,2) + h*h);
constraint S <= PI;maximize V;
}
https://www.localsolver.com/docs/last/exampletour/optimalbucket.html
14 40
Facility location
Select a subset H among N points minimizing the sum of distances from each point in N to the nearest point in H
function model() {
x[1..N] <- bool() ; // decisions: point i belongs to H if x[i] = 1
constraint sum[i in 1..N]( x[i] ) == H ; // constraint: H points selected among N
minDist[i in 1..N] <- min[j in 1..N]( x[j] ? Dist[i][j] : InfiniteDist ) ; // expressions: distance to the nearest point in H
minimize sum[i in 1..N]( minDist[i] ) ; // objective: minimize the sum of distances
}
https://www.localsolver.com/docs/last/exampletour/pmedian.html
15 40
Mathematical operators
+ operator call : to call an external native functionwhich can be used to implement your own (black-box) operator
Decisional Arithmetical Logical Relational Set & List
bool sum sub prod not == count
float min max abs and != contains
int div mod sqrt or at
set log exp pow xor indexOf
list cos sin tan iif > disjoint
floor ceil round array + at < partition
dist scalar piecewise
+ multiobjective capabilities
16 40
Traveling salesman
function model() {
x <- list(N) ; // order the N cities {0, ..., N-1} to visit
constraint count(x) == N; // exactly N cities to visit
minimize sum[i in 1..N-1]( distance( x[i-1], x[i] ) ) + distance( x[N-1], x[0] ); // minimize traveled distance
}
Could you imagine a simpler model• Textbook-like, natural declarative model
• Compact, highly-scalable
Find the shortest tour that visits N cities exactly once
https://www.localsolver.com/docs/last/exampletour/tsp.html
17 40
Vehicle routing
function model() {
routes[1..K] <- list(N) ; // for each vehicle, the list of visited clients
constraint partition[k in 1..K]( routes[k] ); // each client is visited once
for [k in 1..K] {route <- routes[k]; n <- count( route );
constraint sum( 0..n-1, i => demands[ route[i] ] ) <= capacity; // truck capacity constraint
distances[k] <- sum( 0..n-2, i => distance( route[i], route[i+1] ) ) // sum of distances between each pair+ distance( depot, route[0] ) + distance( route[n-1], depot ); // of clients along the route
}
minimize sum[k in 1..K]( count(routes[k] > 0 ); // minimize the number of trucks usedminimize sum[k in 1..K]( distances[k] ); // minimize sum of traveled distances
}
Find the shortest routes of a fleet of K vehicles with a given capacity to deliver N customers
https://www.localsolver.com/docs/last/exampletour/vrp.html
18 40
LocalSolver modeling practices
1. Choose the right type and set of decisions
2. Do not limit yourself to linear operators
3. Model business constraints as primary objectives
4. Think of using external functions
5. Experiment reformulations
Lost in modeling? [email protected]
19 40
LocalSolverBenchmarks
20 40
Quadratic assignment
Assign facilities to locations with quadratic costs• Instances coming from QAPLIB
• Resolution time: 1 minute
• Main difficulty: quadratic expressions to manage in the objective
1 minute
x = no solution found
21 40
Bin packing
Pack items into bins• Instances coming from BPP library (Falkenauer et al.)
• Resolution time: 1 minute
• Main difficulty: standard formulations induced a lot of symmetries
1 minute
22 40
Traveling salesman
Find the shortest tour that visits all cities once• Instances coming from TSP library
• Resolution time: 1 minute
• Main difficulty: specific branch-and-cut required to get decent results
1 minute
23 40
Traveling salesman
Find the shortest tour that visits all cities once• Instances coming from TSP library
• Resolution time: 1 hour
• Main difficulty: specific branch-and-cut required to get decent results
•
1 hour
24 40
Vehicle routing
Find the shortest routes of a fleet of vehicles • Instances coming from CVRPLIB
• Resolution time: 1 minute
• Main difficulty: no general-purpose solver available to get decent results
• Many declinations to manage in real life: CVRPTW, PDP, DARP, LRP, …
1 minute
25 40
Car sequencing • Smoothing car production loads along the assembly line
• 2005 ROADEF Challenge http://challenge.roadef.org/2005/en
• Up to 1,300 vehicles to sequence → 400,000 binary decisions
Instance with 540 vehicles• Small instance: 80,000 variables including 44,000 binaries
• State of the art: 3,109 (found by winner of the Challenge)
• Lower bound: 3,103
Benchmarks• MIP solver 3,027,000 in 10 min 194,161 in 1 hour
• LocalSolver 3,476 in 10 sec 3,114 in 10 min
Industrial benchmark
Minimization Lower is better
26 40
LocalSolverBusiness cases
27 40
100 clients in 20 countries
100 clients
Especially major industrial corporations in Europe
Airbus, Air Liquide, Beiersdorf, Bosch, CEZ, Dassault Aviation,
Deutsche Post, EDF, ENGIE, Norsk Hydro, Primagaz, PSA,
Renault, SNCF, Siemens, SITA, Tetra Pak, Thales, Veolia
Production schedulingWarehouse optimization
Vehicle routingWorkforce planning
Maintenance planning
28 40
Pasco Japan supply chain optimization
Problems involving tens of millions of variables
Solved in minutes thanks to LocalSolver
Daily used since 2015
Testimonial by Mr. Shinichi Kuroda, Pasco Shikishima
29 40
Optimization of Starbucks’ supply chain network
= Location Routing Problem (LRP)
Easily solved using LocalSolver• Huge problem: 200 depots, 20,000 stores
• Easy use of Python API
• List variables make the model much simpler than using a traditional MIP boolean modeling approach
Please have a look at the testimonial by Dr. Renaud Lecoeuche, Principal Data Scientist from Starbucks, USA, in the brochure
30 40
LNG supply chain optimizationLong-term planning of delivery routes and storagesStakes in €bn
31 40
32 40
Assigning trains to platforms
while respecting crossing constraints
33 40
Optimizing rotations of locomotives & drivers for the freight division of the French Railways
34 40
Optimization of the routing & scheduling of field service technicians at JCDecaux France
35 40
Building several networks simultaneously
Considering Points Of Interest
User can modify the solution manually
Selection optimizes• Number of contacts• Number of panels• Geographical dispatch• Mandatory or forbidden panels• Language balancing • And so on…
Programmatic advertising for out-of-home (OOH) media campaigns at JCDecaux Belgium
39 40
High-quality support
All-inclusive maintenance & support• Included in the license price
• Provided by PhD-level optimization experts
• Dedicated: email, phone, web or physical meeting
• Reactive: we answer you during the day, generally within hours
• We help you to model & solve your problems using LocalSolver
• New versions for free (two per year)
• Corrective patches for any version
• Free migration of your license in case of replacement of your hardware
40 40
Thanks
This is the first coming of the LocalSolver team in Japan
We are so pleased to come here in TokyoWarm welcome to all of you for joining this meeting
Special thanks to our Japanese partner MSI for the organization
Frédéric GardiCo-Founder & Managing Partner
Tiphaine RougerieOptimization Engineer
41 40
LocalSolverMore business cases
42 40
Network optimization
ADSL network expansion planning
Context• Choose remote concentrator units to unbundle
• Local and global constraints. Ex: forbid paths with too much clients to limit impacts of an incident
Prize collecting Steiner forest problem• Network: 14,000 nodes, 180,000 edges
• Resulting model: 1.4 million variables
• Required resolution time: minutes
43 40
Energy management
Hydro power plant production optimization• Hydroelectric dams with pumps
• Forecasted energy prices over the horizon
• Management of thermal power plants
• From daily to yearly horizon
→Nonlinear large-scale dynamic system with mixed-variable (0/1 + continuous) decisions and tight coupling constraints
Outperform MILP approaches on the hardest instances
44 40
Nuclear power plant maintenance
Quadratic assignment problem• Placement of nuclear fuel assemblies in pools
• Minimizing the completion time of the operations
• Inducing millions of variables
• Out of scope of MILP & MIQP solvers
Testimonial by EDF R&DLocalSolver follows the claims made by its designers. It was able to adaptand to provide good-quality solutions to the problem of placement ofnuclear fuel assemblies in pools in very short running times on a standardcomputer. It was able to outperform a simulated annealing algorithmwhich however considered the structure of the problem.
45 40
Infrastructure maintenance planning
Street lighting maintenance planning
« Illuminate better with less energy » • Plan the replacement and maintenance
of street lighting fixtures over 25 years
• Considering costs, resources, energy consumptions, failure rates, etc.
Large-scale combinatorial optimization• Thousands of streets, tens of thousands of lighting fixtures
• Nonlinear discrete model with 1,000,000 variables
• Resolution time: minutes
46 40
Oil refinery optimization
Software solutions based on LocalSolver• Refinery planning solution
• Refinery scheduling solution
• Product blending solution
• Crude blending solution
Why moving from MILP solvers? • Broader modeling scope + ability to
connect external simulation codes
→ simpler resolution approaches
• Faster & more scalable
• Better, expert & friendly, support
• Competitive pricing agreement
47 40
Banking & finance optimization
How combining mortgages at best to satisfy a financial need?
Example of plan composed of two mortgages• Financial need €300,000, monthly payment €1,200
• Total cost €415,000 -> gain by composition €15,000
€200,000
240 months
2,20%
€100,000
331 months
2,58%
High-quality solutions required in seconds by bank branch sales
48 40
Media planningSelling advertising spaces in Paris underground
60,000 ad faces to partition into products from 100 to 500 faces• Covering a maximum of stations
• Balanced according to the traffic
• + a dozen of quality criteria
Huge partitioning problemsolved in 1 minute
49 40
Mechanical system design
Designing sailboat weathervanes
Context• Used to measure the wind, to drive sailboats
• High precision needed in racing competition
Heterogeneous variables• Continuous decisions
• Highly nonlinear physics
• Precision: 0.1 millimeter
• 2 criteria optimized
50 40
Agronomic optimization
Optimal fertilization of agricultural parcels
How to best fertilize soils from mineral and organic fertilizers?
Highly-nonlinear dynamic system• Nonlinear dynamics of N, C, K over time
• Thousands of 0/1 and continuous decisions
• Analytical or simulation-based
51 40
Engine calibration
Calibrate engine cartography• Nonlinear parametric models
• Large scale: 5,000 parameters
• Plugged to Matlab Simulink statistical models: 10 calls per sec
• 1st objective: fitting experimental measures
• 2nd objective: smoothing the cartography (avoiding large variations)
• LocalSolver allows to reduce computation time from weeks to hours
52 40
Data science applications
Sparse Least Square Linear Regression
Find a sparse fitting model with only a few nonzero parameters
Applications in many fields• Data analysis in geomarketing
• Data analysis in email marketing
• Data analysis in biotechnologies
min𝛽
𝑌 − 𝑋𝛽 2
𝑠𝑡. 𝛽 0 ≤ 𝑘
53 40
Data science applications
Map antennas in a hidden telco network
ContextHaving approximate distances between antennas from signals, find the positions of the antennas in the network
Unconstrained quadratic problem• 25 antennas, 10,000 point-to-point signals
• Resulting model: 100,000 float variables
• Resolution time: 1 minute
www.localsolver.com