Machine Learning & AI
in Transport and Logistics
Frank Salliau & Sven Verstrepen
Logistics Meets Innovation
Vlerick Brussels – Nov. 15th 2017
Sci-fi in 2002
Reality in 2017
Sci-fi in 1984
Reality in 2017
“Alexa, check my calendar”
“Alexa, what is machine learning?
“Alexa, play Spotify”
What is it all about?
What is Machine Learning?
“The science of getting computers to act
without being explicitly programmed” -
Andrew Ng (Stanford/DeepLearning AI)
Challenge
Recognize dogs in images
What a human sees
DOG NOT A DOG
What a computer sees
??? ???
What do we need?
TRAINING DATA
Training phase
DOG (1)
NOT DOG (0)
Labeled training set (dog/not dog) Untrained Neural Network
> 1000 images
Prediction phase
87% DOG
13% NOT DOG
Unlabeled image Trained Neural Network
Why is this booming now?
What are the drivers?
(BIG) DATA
Byte : one grain of rice
Byte
Credits: David Wellman, http://www.slideshare.net/dwellman/what-is-big-data-24401517
Byte : one grain of rice
Kilobyte : cup of rice
Kilobyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Megabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Gigabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Terabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Petabyte : Blankets Manhattan
Petabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Petabyte : Blankets Manhattan
Exabyte : Blankets west coast states
One Byte
Exabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Petabyte : Blankets Manhattan
Exabyte : Blankets west coast states
Zettabyte : Fills the Pacific Ocean Zettabyte
Byte : one grain of rice
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Petabyte : Blankets Manhattan
Exabyte : Blankets west coast states
Zettabyte : Fills the Pacific Ocean Yottabyte
Yottabyte : A EARTH SIZE RICE BALL!
Byte : one grain of rice
Hobbyist
Kilobyte : cup of rice
Megabyte : 8 bags of rice
Desktop
Gigabyte : 3 Semi trucks
Terabyte : 2 Container Ships
Internet
Petabyte : Blankets Manhattan
Exabyte : Blankets west coast states
Big Data
Zettabyte : Fills the Pacific Ocean
Yottabyte : A EARTH SIZE RICE BALL! The Future
1 Yottabyte
1 Xenottabyte
1 Shilentnobyte
1 Domegemegrottebyte
1 Icosebyte
1 Monoicosebyte
Where does all this data
come from?
The Power of the Crowd
Machine Learning
Supervised Unsupervised Reinforcement
Learning Learning Learning
Deep Learning Deep
Reinforcement
Learning
Supervised Learning
Regression
Classification
Deep Learning
Google Deep Dream
(sometimes nightmare)
Unsupervised Learning
Reinforcement Learning
Getting Started
CRISP Methodology
Exploratory
Iterative
Quality of data: Garbage In Garbage Out
Multi-disciplinary team
D3.js - Open Source
Examples in Transport & Logistics
IBM Watson
Ahlers Supply Network Innovation & Analytics (ASNIA)
Using ML to identify horizontal collaboration synergies between multiple shipper networks
Transmetrics: replacing budgets with prediction
Transmetrics big data cargo platform – a rigorous approach deriving benefits from current and future data
1 2 3 4
Data uptake Demand AI Execution
cleansing forecast optimizatio Controlling
and modelling n
enrichment
Traditional tactical management relies on budgets Transmetrics: predictive tactical management
Budgeting process: Transmetrics:
“Flying on paper maps, experience and gut feeling” “High precision flying assisted by data and AI”
Data quality improvement – status of data
after AUTOMATED / AI processing
Preliminary data quality assessment after AUTOMATED / AI processing
Challenging Good to
very good
Original situation Achieved situation
• Ambiguity in interpretation of • Clear and correct identification of all shipments
orders and pieces records and pieces (including ignoring of non-piece lines)
• Missing size information at • For each piece calculated full set of
order and piece level measurements (height, width, length, weight,
volume) – which enables 3D loading factors!
• Mix of measurement units (ldm, • Complete set of measurements: loading meters,
m3, kilos, pallets) surface/pallets (m2), volume (m3)
• Missing or unreliable capacity • Built AI algorithm to estimate missing capacity
information for linehauls information
• Other issues (e.g. senders with • Grouped senders by AI; among others found a
multiple name spellings) customer with 330+ different accounts/names
Methodology / Behind the scenes
1. Preprocess 2. Estimate 3. Estimate dimensions
data volume • Mine text fields,
• cleanse and disaggregate • Predict missing • Apply industry standards and
pieces/volume/weight volume data business rules
• Remove unnecessary rows
• Ensure one row=one piece
Approaches used 4. Predict missing
linehaul capacity
• Combinatorial
optimization • use historical data to
predict capacities
• Quadratic
optimization • 95% accuracy
• Gradient boosting trees
and other regressions 3D Loading Factor estimation
• Natural language • Loading meters (m)
processing • Loading area, pallets (m2)
• Named identity • Volume (m3)
disambiguation
• Expected 90% accuracy
• Expert input / after calibration
business rules
3D loading visualization
Loading factor optimization results
Real situation
(as observed on
the warehouse
floor)
Loading factor Loading Floor Volume
calculated by meters
using the data 13.6 m 30.4 m2 29.8 m3
100% 91% 34%
Sven Verstrepen
Head of Supply Network Innovation & Analytics
[email protected]
http://www.linkedin.com/in/svenverstrepen
http://www.mentat-it.be
Frank Salliau
Independent Data Scientist & Machine Learning Expert
[email protected]
https://be.linkedin.com/in/franksalliau
