How Zalando Accelerates Warehouse Operations with Neural

Networks

Calvin SewardBig Data Berlin v. 6.0

28 January 2016

1 / 22

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

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

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

Picker Routing ProblemThe locations that must be visited

4 / 22

Picker Routing ProblemThe locations that must be visited

5 / 22

Picker Routing ProblemThe Pick Route

6 / 22

Picker Routing ProblemThe Pick Route

7 / 22

Picker Routing ProblemThe Pick Route

8 / 22

Picker Routing ProblemThe Pick Route

9 / 22

Picker Routing ProblemThe Pick Route

10 / 22

Picker Routing ProblemThe Pick Route

11 / 22

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

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

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

Simpli�ed Order Batching ProblemBipartite Graph Formulation

n Orders

o1

o2

o3

. . .

on

m Pick Tours

t1

t2

. . .

tm

13 / 22

Simpli�ed Order Batching ProblemBipartite Graph Formulation

n Orders

o1

o2

o3

. . .

on

m Pick Tours

t1

t2

. . .

tm

13 / 22

Simpli�ed Order Batching ProblemBipartite Graph Formulation

n Orders

o1

o2

o3

. . .

on

m Pick Tours

t1

t2

. . .

tm

13 / 22

Order Batching ProblemRandom Split of 10 Orders à 2 Items Into Two Pick Routes

14 / 22

Order Batching ProblemBrute Force Split of 10 Orders à 2 Items into Optimal Two Pick Routes → 8.3% Boost

15 / 22

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

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

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

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

Neural Network Estimate of Pick Route LengthEstimation Accuracy � Frequency of relative estimation error estimated travel time

calculated travel time

18 / 22

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

Order Batch optimization via Simulated AnnealingEstimated and Exact Improvement in Example Simulated Annealing Run

20 / 22

Questions?

21 / 22

Thanks for listening

Get The Slides Ongithub.com/cseward/ocapi_neural_net_blog_post

22 / 22