This document describes two use cases for route optimization using artificial intelligence: last mile delivery route optimization and long distance route optimization. For last mile route optimization, the problem was a wine distributor's manually planned delivery routes, which lacked digitalization. The solution used an algorithm to optimize routes based on vehicle capacity, time windows, and other data. This reduced planning time by 25% and routes by 90%. For long distance route optimization between warehouses, the complexity was reduced by dividing the problem and finding repetition patterns. This included 75% of routes repeating weekly and restricted empty vehicle returns. Next steps discuss applying multi-agent reinforcement learning.
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
From last mile to long distance route optimization: Intelligent Mobility solutions for logistics
1. From last mile to long distance route optimization
Intelligent Mobility
2. Table of
Content
INTRODUCTION
Uses cases overview
Who we are
USE CASE 1:
Solution
Problem
Results
Last mile route
optimization
USE CASE 2:
Results
Problem
Solution
Long distance routes
optimization
02
From last mile to long distance
route optimization
NEXT STEPS
4. Who
we are
Hedyla is a software company
dedicated to optimize logistic
processes through AI
04
5. Use cases overview
05
Last mile Long distance
Reduce the number of
vehicles, distance and time
Common use case
AI algorithms can be applied directly
Main goal
Key data
Not common
Combination of different AI algorithms
Current status
Reduce the distance
travelled with no load
Vehicle capacity, delivery and driver time windows,
traffic, business restrictions
Loads exchangeability, exchange locations,
drivers origin and destination
Intuitive solution Find complementary routes
Group deliveries that are close to the others
Main complexity
Detailed data accuracy:
traffic, time windows...
High data volume
7. 07
Initial routes created by a traffic responsible, then drivers themselves
exchanged deliveries based on their experience and knowledge.
Daily routes created manually
Most of the information not digitalized. Critical information like client
preferred delivery time windows known only by the drivers, a lot of
dependency on their experience.
Information not centralized
Impossible to optimize other processes
Due to lack of digitalization and uncertainty of the process, there were
other processes, like the preparation in the warehouse, that could not be
optimized.
Problem
1
2
3
Wine distributor.
Delivery to shops and e-commerce
with 500-800 deliveries per day.
2 shifts: morning and afternoon.
~25 vehicles between trucks and
vans.
Optimize route planning to
reduce costs.
8. Involve key people
It's critical to find benefits for each actor:
drivers and management.
User interface simplicity
The complexity must be hidden. Most of the
parametrization derived from a short list of
basic options.
Adaptative algorithm
The algorithm must work in different
situations. Resources availability and demand
are variable.
Provide understandable results
Be able to explain the optimization results
(high level). Less optimal solution with an
obvious explanation is better than a very
optimized result, with no clue.
08
Solution Keys to success
Vehicles (capacity)
Drivers working hours.
Clients geolocation and delivery
time windows.
Stop time: fixed time by client +
variable time based on the weight
of the load.
Data preparation
1
Define goals: time optimization +
improve quality of service.
Define restrictions: Common + drivers
by area
Algorithm: complex VRP solver based
on parallel insertion strategy.
Extra logic: Orders preparation
optimization in the warehouse
Business logic
2
Bidirectional integration of clients,
orders and deliveries status.
All information centralized for
analytics and other applications.
Integration
3
Traffic responsible: Routes plan
process
Drivers: Deliveries tracking from mobile
app.
Training
4
9. 09
less time spent
executing routes
25% Planning time saved
90%
Deliveries digitalized
Information
centralized
Prepared for
next challenges
Results
Areas of improvement
Automatically adjust stop
times based on historic data
Suggest changes on client delivery
time windows based on incidences
1 2
11. 11
Each warehouse is responsible of their provisioning and they have a limited
vision of the rest of routes. They try to find combinations by talking to other
warehouses and providers, but it's a manual and long process.
Routes are managed by each warehouse
Multiple restrictions: different type of loads, vehicles with different
temperatures, drivers working time limitations, limited number of docks per
warehouse, pickup and delivery time windows...
High complexity
The amount of data is huge
Although all data is centralized and there are departments with overall
vision, there is too much data and restrictions to handle it properly with
traditional strategies.
Problem
1
2
3
Supermarket chain.
~800 providers and ~20 warehouses
around Spain.
~2.000 loads per day and ~5.000
trucks.
Trucks must come back to the origin
every day.
Routes are long, so trucks exchange
loads very often. The same load can be
carried by two or three different trucks
while travelling from origin to
destination.
Routes must be repeatable in different
days and weeks.
Client can't be disclosed
Find repeatable combined long
distance routes to minimize the
empty return travels.
12. Divide et Impera
It was impossible to solve the problem
globally, so we divided the problem in
independent blocks to find optimal solutions
in parallel.
12
Solution Keys to success
Main goal: reduce the search space:
Find repetition patterns in the orders among days of the week and
weeks in a month.
Find spatial patterns to create routing hierarchy.
Extensive data analysis
1
Solution design and implementation
2
Multiple steps and techniques
As the problem was divided, there were
different subproblems, so different techniques
were applied to optimize each subgroup of
orders, depending on the nature of the
problem.
AI general knowledge
In these complex cases, a generalist approach
will never work. Understanding the limitations
and possibilities of different algorithms is
critical to find feasible solutions.
Skip not repeated loads, can't be used to generate repeatable combined
routes.
Generate subgroups of loads by frequency and compatibility in order to
divide the problem.
Generate intermediate locations to allow loads exchanges between trucks.
Apply VRP solver to subproblems.
13. 13
of orders included in
a repetitive route.
75% routes repeat every week
at least for a month
800
All orders included Restrictions
low vehicles
capacity usage
Results
15. 15
Next steps RL is a model based on a reinforced learning principle. It represents the
state of the art in artificial intelligence.
Learn to learn
Each agent has a state that represents its load, driving time and can
choose between different primitive actions: load, unload, travel...
Trucks are the agents
Agents are rewarded based on results
Positive reward when constraints are met, and negative otherwise.
Distributed logic system
Each agent looks for its own optimization
Investigate the
application of an
alternative method:
Multi-Agent
Reinforcement
Learning (MARL)