how zalando accelerates warehouse operations with neural networks - calvin seward, data scientist at...
TRANSCRIPT
![Page 1: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/1.jpg)
How Zalando Accelerates Warehouse Operations with Neural
Networks
Calvin SewardBig Data Berlin v. 6.0
28 January 2016
1 / 22
![Page 2: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/2.jpg)
Outline
I Picker Routing Problem
I Order Batching Problem
I Neural Network Estimate of Pick Route Length
I Order Batch optimization via Simulated Annealing
This was a project that was a collaboration between Rolland Vollgraf, Sebastian Heinzand myself. Some of the �gures have been shamelessly stolen from Roland.
2 / 22
![Page 3: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/3.jpg)
Outline
I Picker Routing Problem
I Order Batching Problem
I Neural Network Estimate of Pick Route Length
I Order Batch optimization via Simulated Annealing
This was a project that was a collaboration between Rolland Vollgraf, Sebastian Heinzand myself. Some of the �gures have been shamelessly stolen from Roland.
2 / 22
![Page 4: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/4.jpg)
Zalando's Logistics Centers
I 13.8 Million orders from 01.07.15 � 30.09.15
I 174,684 orders send per working day (Mo. � Sa.)
I Every second handling time per order requires 160 man-days of work / month
I Any increase in e�cency has a big impact
3 / 22
![Page 5: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/5.jpg)
Picker Routing ProblemThe locations that must be visited
4 / 22
![Page 6: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/6.jpg)
Picker Routing ProblemThe locations that must be visited
5 / 22
![Page 7: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/7.jpg)
Picker Routing ProblemThe Pick Route
6 / 22
![Page 8: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/8.jpg)
Picker Routing ProblemThe Pick Route
7 / 22
![Page 9: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/9.jpg)
Picker Routing ProblemThe Pick Route
8 / 22
![Page 10: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/10.jpg)
Picker Routing ProblemThe Pick Route
9 / 22
![Page 11: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/11.jpg)
Picker Routing ProblemThe Pick Route
10 / 22
![Page 12: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/12.jpg)
Picker Routing ProblemThe Pick Route
11 / 22
![Page 13: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/13.jpg)
OCaPi AlgorithmOptimal CArt PIck
I To solve picker routing problem, wedeveloped the OCaPi Algorithm
I Calculates the optimal route to walk
I Also determines optimal cart handlingstrategy
I Has complexity that is linear in thenumber of aisles
I Unfortunately still has a runtime ofaround 1 second
Figure: The Okapi � Our Mascot
12 / 22
![Page 14: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/14.jpg)
OCaPi AlgorithmOptimal CArt PIck
I To solve picker routing problem, wedeveloped the OCaPi Algorithm
I Calculates the optimal route to walk
I Also determines optimal cart handlingstrategy
I Has complexity that is linear in thenumber of aisles
I Unfortunately still has a runtime ofaround 1 second
Figure: The Okapi � Our Mascot
12 / 22
![Page 15: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/15.jpg)
OCaPi AlgorithmOptimal CArt PIck
I To solve picker routing problem, wedeveloped the OCaPi Algorithm
I Calculates the optimal route to walk
I Also determines optimal cart handlingstrategy
I Has complexity that is linear in thenumber of aisles
I Unfortunately still has a runtime ofaround 1 second Figure: The Okapi � Our Mascot
12 / 22
![Page 16: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/16.jpg)
Simpli�ed Order Batching ProblemBipartite Graph Formulation
n Orders
o1
o2
o3
. . .
on
m Pick Tours
t1
t2
. . .
tm
13 / 22
![Page 17: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/17.jpg)
Simpli�ed Order Batching ProblemBipartite Graph Formulation
n Orders
o1
o2
o3
. . .
on
m Pick Tours
t1
t2
. . .
tm
13 / 22
![Page 18: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/18.jpg)
Simpli�ed Order Batching ProblemBipartite Graph Formulation
n Orders
o1
o2
o3
. . .
on
m Pick Tours
t1
t2
. . .
tm
13 / 22
![Page 19: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/19.jpg)
Order Batching ProblemRandom Split of 10 Orders à 2 Items Into Two Pick Routes
14 / 22
![Page 20: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/20.jpg)
Order Batching ProblemBrute Force Split of 10 Orders à 2 Items into Optimal Two Pick Routes → 8.3% Boost
15 / 22
![Page 21: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/21.jpg)
Neural Network Estimate of Pick Route Length
I This simple example could be donewith brute force
I A realistic example with 40 orders à 2items has a complexity of
40!
2 · 20! · 20!≈ 6.9 · 1010
at 1 second per route, you'd wait 2185years
I Use clever heuristics like simulatedannealing
I Estimate pick route length with NeuralNetworks
16 / 22
![Page 22: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/22.jpg)
Neural Network Estimate of Pick Route Length
I This simple example could be donewith brute force
I A realistic example with 40 orders à 2items has a complexity of
40!
2 · 20! · 20!≈ 6.9 · 1010
at 1 second per route, you'd wait 2185years
I Use clever heuristics like simulatedannealing
I Estimate pick route length with NeuralNetworks
16 / 22
![Page 23: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/23.jpg)
Neural Network Estimate of Pick Route Length
I OCaPi cost landscape
f : (N× R)n → R+
is a nice function because it is:I Lipschitz continuous in the real-valued
argumentsI Piecewise linear in the real-valued argumentsI Locally sensitive
I Perfect function to model with ConvolutionalNeural Networks with ReLUs:
f̃ (x) := (W2(W1x + b1)+ + b2)+
I Train convolutional neural network with 1million examples
17 / 22
![Page 24: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/24.jpg)
Neural Network Estimate of Pick Route Length
I OCaPi cost landscape
f : (N× R)n → R+
is a nice function because it is:I Lipschitz continuous in the real-valued
argumentsI Piecewise linear in the real-valued argumentsI Locally sensitive
I Perfect function to model with ConvolutionalNeural Networks with ReLUs:
f̃ (x) := (W2(W1x + b1)+ + b2)+
I Train convolutional neural network with 1million examples
17 / 22
![Page 25: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/25.jpg)
Neural Network Estimate of Pick Route LengthEstimation Accuracy � Frequency of relative estimation error estimated travel time
calculated travel time
18 / 22
![Page 26: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/26.jpg)
Neural Network Estimate of Pick Route LengthEstimation Speed � Time Per Route on two Intel Xeon E5-2640 and two NVIDIA Tesla K80 accelerators
number pick lists OCaPi CPU network GPU network
1 5.369 2.202e-3 1.656e-310 1.326 1.991e-4 1.832e-4100 0.365 6.548e-5 5.919e-51000 3.086e-5 2.961e-510000 2.554e-5 2.336e-5
19 / 22
![Page 27: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/27.jpg)
Order Batch optimization via Simulated AnnealingEstimated and Exact Improvement in Example Simulated Annealing Run
20 / 22
![Page 28: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/28.jpg)
Questions?
21 / 22
![Page 29: How Zalando accelerates warehouse operations with neural networks - Calvin Seward, Data Scientist at Zalando](https://reader034.vdocuments.site/reader034/viewer/2022052205/58713aff1a28abf0568b6cef/html5/thumbnails/29.jpg)
Thanks for listening
Get The Slides Ongithub.com/cseward/ocapi_neural_net_blog_post
22 / 22