2015 Artificial Intelligence Techniques at Engineering Seminar - Chapter 2 - Part 4: Intelligent Transportation Systems

Artificial Intelligence Techniques
applied to Engineering
Part 2. Genetic Fuzzy Systems
Enrique Onieva Caracuel
@EnriqueOnieva
1.Fuzzy Logic
2.Genetic Algorithms
3.Genetic Fuzzy Systems
4.Applications to Intelligent Transportation
Systems: My Experience

Técnicas de Inteligencia Artificial aplicadas a la Ingeniería
@EnriqueOnieva 2014/2015 2
Intelligent Transportation Systems
Intelligent Transportation Systems integrate
information and communication technologies with
transportation of passengers and goods
 Mobility
 Safety
 Productivity
 Energy consumption
 Capacity
1

Intelligent Transportation Systems
 Common services
 Information Systems
 Route planning
 Air transport
 Maritime transport
 Road transport
 Intelligent infrastructure
 Intelligent vehicles
Active assistances
Pasive assistances
2
Autonomous Driving?

Motivation
 Fuzzy Logic
 IF the vehicle is derived through the right, steer to the left
 IF the vehicle is derived through the left, steer to the right
 IF the vehicle is slow, press the throttle
 IF the vehicle is fast, press the brake
1
Control System Driver

Motivation 2

AUTOPIA Program 1
1998
2012

AUTOPIA Program
 Throttle
 Pedal signals commuted
 Orders communicated by an
Analog Card
 Brake
 Intervention on the ABS
 Electro-hydraulic system
Motor
Deposit
3 valves: Limiter, Proportional,
Nothing-All
 Orders communicated by a
CAN controller
2
WLAN
Antenna
Power
Supply
GPS
Receiver
Computer
IMU
GPS
Antenna
CAN-USB
converter
Auxiliary
Battery
CAN
Module
Electro-
hydraulic
system

Speed Control for Cities
 Inputs
 Speed error(km/h)
 Acceleration (km/h/s)
 Outputs
 Throttle [0,1] Throttle [0,0.4]
 Brake [0,1] Brake [0,0.2]
1
Comfort
acceleration
≤ 2.5 m/s2
Ac+ Ac0 Ac-
EV+ B02 B01 B01
EV0 T00 t01 T01
EV- T01 T02 t04
E. Onieva, et al., Throttle and Brake Pedals Automation for Populated Areas, Robotica, vol. 28, n. 4, pp 509-516.
Speed Error (km/h)
Acceleration (km/h/s)
Negative NegativeZero Positive
Negative Zero Positive
T00 T01 T02 T04
B00 B01 B02

Speed Control for Cities
 First gear (speed > 16 km/h)
 Error measured after transitory
state (5 s)
 Bigger error at 15 km/h
 (1st to 2nd gear)
 Similar results (speed≤ 20 km/h)
 Better results at 25 km/h
2
E. Onieva, et al., Throttle and Brake Pedals Automation for Populated Areas, Robotica, vol. 28, n. 4, pp 509-516.
Human System
10km/h ±0.63 ±0.71
15km/h ±0.88 ±0.98
20km/h ±0.72 ±0.84
25km/h ±1.22 ±0.9
Speed Target
Speed Target
Speed Target
Speed Target
Speed Target
Time (s)
Speed(km/h)

Speed Control on-line Learning
1. Rules’ consequents modification in real time
Speed error
Acceleration
Rule activation
 9 cased reward
Positive or Negative error
Acceleration and comfort acceleration
Acceleration decreases when error  0
1
E. Onieva et al., On-Line Learning of a Fuzzy Controller for a Precise Vehicle Cruise Control System, Expert Systems with Applications. 40 (4) , pp. 1046-1053.

1. Rules’ consequents modification in real time
2
Speed(km/h)
Time (s)

2. Trapezoids’ modification
 After a certain time (100 seconds)
 Input values histogram analysis
A trapezoid is added if it is low-covered
A trapezoid is narrowed if covers several frequent values
3

Speed Control on-line Learning 4
 Test with 40 vehicles in TORCS
 Different dynamics
 Different behaviors
 Simple initial controller
 2x2 membership functions
 All the singletons = 0 (do nothing)

1. Speed change test
2. Fixed speed test (15 km/h)
3. Fixed speed test (5 km/h)
5

Information capture
Information Processing
Simplification
Extension
Steering control by Genetic
Algorithms 1
E. Onieva, et al., Genetic fuzzy-based steering wheel controller using a mass-produced car, Int. J. of Innovative Computing, Information and Control. vol. 8, n. 5B, pp. 77-94.
E. Onieva, et al., Automatic lateral control for unmanned vehicles via genetic algorithms. Applied Soft Computing Journal. 11 - 1, pp. 1303 - 1309, 2011
Lateral Error Angular Error
Steering
Lateral Error Angular Error
Steering
Lateral
Error
Angular
Error
Reference Line

Algorithms
 Membership functions Representation
 Rule base representation
 Integer coding
 Length = Number of rules
 21 Singletons in [-1,1]
2
LAT/ ANG
θ
{VeryLeft}
Left
No
Right
{VeryRigth}
Θ M M M M M
{VeryLeft} M C C C C C
Left M C C C C C
No M C C C C C
Right M C C C C C
{VeryRigth} M C C C C C
Left NO Right VeryLeft VeryRightLeft NO Right
R10, R9, R8, R7, … R1 L1, L2, L3, L4, … L10NO
Right (Clockwise) Left

Algorithms
 Genetic fuzzy system
in 2 stages
 Membership function
optimization
Real coding
BLX-𝛼 crossover
 Rule base optimization
Integer coding
One point crossover
 Steady state Genetic Algorithm
 Binary Tournament
 Uniform Mutation
 Worse individual replacement
3

Algorithms
Objective function
 Mean squared error (MSE)
 Highest jump in the control surface (Dist)
Fitness Function (Min): 0.75·MSE + 0.25·Dist
4

Algorithms 5
Controllers Speeds
Labels Rule Base Average Maximum
3x3 Marginal 12.8 22.4
3x3 Central 14.6 22.1
3x3 Total 13.5 24
5x5 Marginal 14.9 22.1
5x5 Central 14.8 28.6
5x5 Total 14.8 27.9
Lateral errorAngular error Honorable
Mention to
best student
work
ESTYLF 2008
East (m)
North(m)

Intersection decision by genetic
algorithms
Decision making in non-
cooperative intersections
Non yielding always
strategy
Safe and efficient
maneuver
 Slightly accelerate to pass
before the manual one
 Slightly brake to yield
 Without stopping
1
E. Onieva, et al., Genetic Optimization of a Vehicle Fuzzy Decision System for Intersections, Expert Systems with Applications. 39 (18) , pp. 13148-13157
Accidents at
intersections
Intelligent
Trasnportation Systems
Intersections
Manual Manual Autonomous
Autonomous Autonomous Autonomous
Coordination
Accidents
Roads

algorithms
1. Check if the vehicle is going to cross and by where
 Fuzzy rule based system  3 inputs
 Manually adjusted
2. Decide the autonomous vehicle’s speed to finish the
maneuver
 Without risk
 As soon as possible
 Fuzzy rule based system  4 inputs
 Coded with {2,3,4} membership functions  81
Granularities
 2 types of outputs
 Relative / Absolute Speed
 162 controllers adjusted by a Genetic Algorithm
3. Move the pedals to reach the desires speed
 Vehicle’s longitudinal dynamics model
 Flat surface
2
Manual
Autonomous
Time
Speed

algorithms
 Evaluation in a variable number of
scenarios
 Nsc=1+19·(g/G)
 2 Executions
 Free (EF)  (Keep SA)
What happens if speed does not vary?
 Controlled (EC)
Does the fuzzy system avoid the
collision
 3 possible results
 No collision
 Lateral collision
 Frontal collision
3
No Collision Lateral Collision Frontal Collision
Keep
Speed
up
Slow
Down

algorithms
 Partial fitness depending on:
 Result in free execution
 Result in controlled execution
 How much has been the speed varied
 Fitness function  Minimize the sum of partial fitnesses
4
Description Meaning Partial fitness
No collision (EF=EC=NO) |∫SA
c-∫SA
l|
A lateral collision is avoided (EF=LA & EC=NO)
•|∫SA
c-∫SA
l|, if speeds up
•2.500, if brakes down
A frontal collision is avoided (EF=FR & EC=NO)
•|∫SA
c-∫SA
l|, if brakes down
•2.500, if speeds up
Collision is not avoided (EF≠NO & EC ≠ NO) 5.000
Collision is caused (EF=NO & EC ≠ NO) 10.000

algorithms
 Safety vs Number of rules
 Safety > 90%
 Some relatives are worse that a ‘do nothing’ system
 Absolute ones are safer
 ABS4423, ABS3433, ABS3344 y ABS2442  100%
 Is the safety dependent on the type (absolute / relative) of
controller?
5
Relative Controllers
Absolute Controllers
Number of Rules

algorithms
 Granularities correlation
 Most systems are near the diagonal
 54% vs 46% of structures are safer with a specific model
 Safe structures are both when relative and absolute
output
6
Safety for Relative FRBS
SafetyforAbsoluteFRBS
Structures with higher
safety in Absolute
mode (46%)
Structures with higher
safety in Relative mode
(54%)

algorithms
 A ‘stop always’ policy  safety 100 %
 No efficient, neither intelligent
 Fitness function measures the efficiency of the systems
 Lineal relationship
 Safety comes with efficiency
7
Relative Controllers
Absolute Controllers
Safety
FitnessFunction

algorithms
 Both vehicles start at same speed
 Free execution  Frontal collision
 System must brake
 All of them do
 ABS4242 brake less
 REL3344 speeds up once the risk disappear
 Autonomous one starts slightly faster
 Free execution  Lateral collision
 System must speed up
 All of them avoid the collision
 REL3344 does it by speeding up
 Autonomous one starts much faster
 Free execution  No collision
 System must maintain speed
 All of them avoid the collision
 REL3344 varies less the speed
8
Time (s)
Speed
(km/h)
Distance
(m)
Time (s)
Speed
(km/h)
Distance
(m)
Time (s)
Speed
(km/h)
Distance
(m)

Videos
Learning to steer in an autonomous vehicle
Self-Archive
Autonomous Driving Citroën C3 Pluriel
https://www.youtube.com/watch?v=qm-nh7_fJvY
Grand Cooperative Driving Challenge (GCDC) - Technische Universiteit Eindhoven
https://www.youtube.com/watch?v=BprHHm5j_hA
Other Applications
Composition: https://www.youtube.com/user/GrupoAUTOPIA/videos
Autonomous Driving @ High Speed
https://www.youtube.com/watch?v=1zoTg_Pnxbg

2009 Simulated Car Racing
Championship 1
Gears
Change current gear Rule system (RPM)
Use reverse gear Angle + Deviation
Steer
Centered vehicle Laser Sensors
Reverse gear Angle
Go back to track Angle + Deviation
Pedals
Adequate speed Speed error
ABS / TCL Filters Speed-Wheels
Objective Desires Speed Fuzzy System
Learning
Off track
Decision systemBorders crashs
Long straights
Opponents
Overtaking
Rule SystemAvoid collisions
Emergency braking
E. Onieva, et al., A Fuzzy Based Driving Architecture for Non-player Characters in a Car Racing Game, Soft Computing vol. 15, n. 8, pp. 1617-1629, 2011.

Championship
Gear Control
 Change current gear according with RPM
[1ª-3ª] ↑ if RPM>9000
[4ª-5ª] ↑ if RPM>9500
[2ª-4ª] ↓ if RPM<3000
[5ª-6ª] ↓ if RPM<3500
 Reverse gear?
 Continue race
2

Championship
Pedals control
 Throttle and brake
Pedal [-1,1]
 Speed and wheels’ speed based filters
[ABS  Brake]
[TCS  Throttle]
 Special case: Reverse Gear
Pedal = 0.25
3
Speed – Target (km/h)
Pedal

Championship
Steer control
 Reverse gear
 Off-track
 Inside track
4
Steer
Angle (rad)
Angle (rad)
Steer
Deviation (m)

Championship
Objective speed
 IF FRONT is High  200 km/h
 IF FRONT is Medium  175 km/h
 IF FRONT is Low Y MAX10 is High  150 km/h
 IF FRONT is Low Y MAX10 is Medium  125 km/h
 IF FRONT is Low Y MAX10 is Low Y MAX20 is High  100 km/h
 IF FRONT is Low Y MAX10 is Low Y MAX20 is Medium  75 km/h
 IF FRONT is Low Y MAX10 is Low Y MAX20 is Low  50 km/h
 Non-Fuzzy rule:
 IF FRONT = 100  300 km/h
5
FrontMax10Max20
Low Medium High
Low Medium High
Low Medium High
Front = P0

Championship
Opponents
 Modify the steer to overtake
Sensors at {±90º} SI (measure/speed)<Tolerance  (steer+=Increment)
 Modify the steer to avoid collisions
Sensors at {±30º} SI (measure<10)  (steer±=0.25)
 Emergency breaking
Sensors at {±20º} SI (measure<10)  (VELobj *=0.8)
6

Championship
 3 International Conferences
 Rules
 3 unknown tracks
 Classification phase  race alone: 200 seconds
 Race among the 8 classified.
10 races, 10 laps, different starting
F1 punctuation scheme:
 Fastest lap  +2
 Less damage +2
Final score  Median over 10 races
7

Championship 8
CEC GECCO CIG FINAL
Proposal 22 32 29 83
Cobostar 28.5 16.5 30 75
Champ2008 20 23 12.5 55.5
Perez &Saez 16 11 12.5 36.5
Best student
work award
ESTYLF 2010

Championship
Similar architecture
 Punctual modifications in certain modules
 Removing of the fuzzy system in charge of determining
the desired speed
 Optimized Steer and target speed:
Generational Genetic
Algorithm
Controller evaluation
in 4 tracks
Maximize the sum of
distances coverted
1
E. Onieva, et al., An evolutionary tuned driving system for virtual car racing games: The AUTOPIA driver, Int. Journal of Intelligent Systems, vol. 27, n. 3, pp. 217–241, 2012.

Championship
Real optimization
 10 component vector
Steer control
Target speed
 BLX-𝛼 Crossover
 Uniform mutation
2

Championship
Winner in 2010
 System to beat at
2011, 2012 y 2013
Not beaten until now
3
Proposal Muñoz Mr. Racer Polimi
GECCO_1 12 5.5 9 6
GECCO_2 12 8 4 4.5
GECCO_3 10 9 3 5.5
WCCI_1 10 10 4 5
WCCI_2 8 10 3 5
WCCI_3 6 8 2 6
CIG_1 8 4 3 10
CIG_2 12 2 4 6
CIG_3 5 6 8 8
Proposal Muñoz Mr. Racer Polimi
GECCO 34 17 16 6
WCCI 24 28 9 16
CIG 25 12 15 24
TOTAL 83 57 40 46

Racing overtakes
Opponents that oppose to be
overtaken are implemented.
They try to reach the position
of the overtaker
3 types:
 Limited
 Slow
 Complete
Opponent must be overtaken
1
E. Onieva, et al., Overtaking Opponents with Blocking Strategies Using Fuzzy Logic, IEEE Conference on Computational Intelligence in Games, 2010
Limited
Slow
Complete

Racing overtakes
 Fuzzy Rule Based System
 4 inputs:
 Longitudinal distance (Dx)
 Lateral distance (Dy)
 Lateral deviation (DL)
 Time to Collision (TtC)
 2 outputs
 Required lateral position
 Pedal (Emergency braking)
 Manually tuned rule base
 600 potential rules  3·8·5·5 Labels
 Common sense  If the maneuver is finished go back to the center
 Grouping  If lateral distance is long, do not move
 81 rules in the final rule base
2

Racing overtakes
 They were tested:
 The proposal
 Controllers included in TORCS
 Against:
 Slow at 12 different speeds
 Limited at 12 different speeds
 Complete at 12 different speeds
 It is measured:
 % of finished maneuvers
 % of maneuvers without frontal damage
(system’s fault)
 % of maneuvers without lateral damage
(opponent’s fault)
3
Proposal Berniw Bt Inferno Lliaw Olethros Simplix Tita
%S 100 34.4 21.9 37.5 37.5 31.3 25 37.5
%Bf
D 90.6 75 87.5 78.1 78.1 100 100 81.3
%Lf
D 87.5 62.5 96.9 65.6 65.6 93.8 100 65.6
2º Best Work
IEEE-CIG 2010

Videos
Highlight from the Simulated Car Racing Competition at CEC-2009 - Driver by Onieva and Pelta
https://www.youtube.com/watch?v=k5FgzAmJdzs
2010 Simulated Car Racing Championship - First Leg @ GECCO-2010
https://www.youtube.com/watch?v=SXDJMXpiRs0

 We collect traffic data from the
 California Department of Transportation
 About 15 km long
14 (actually more) loops detectors
6 loops detectors which give us flow, speed and density
8 loops detectors which give us only flow
Congestion Prediction 1
Objective:
When congestion is going
to occur here ?
X. Zhang, E. Onieva, et al., Hierarchical Fuzzy Rule Based System Optimized with Genetic Algorithms for Short Term Traffic Congestion Prediction. Transportation Research Part
C: Emerging Technologies. 43, pp. 127-142. 2014

Congestion Prediction
 We group the information in 14 possible input variables
 3 flows in the main highway F1 F2 F3
 3 densities in the main highway D1 D2 D3
 3 speeds in the main highway S1 S2 S3
 2 input flows from the entrances iF1 iF2
 2 output flows from the exits oF1 oF2
2

We define a hierarchical fuzzy system structure to
predict congestions at desired
 Example: 4 input variables with 3 membership
functions per variable
3
The same example with N input variables :
3N rules in the non-hierarchical system
9·(N-1) rules in the hierarchical system
14 variables:
4.782.969 Vs 117 Rules

The systems is optimized by a Genetic Algorithm:
 3-part coding
Input variables’ order  Variable selection
Membership Functions
Rules’ consequents
 2 operator groups:
Permutation
Real Coding
4

3 experiments:
 97% 5 minutes ahead 9 variables
5

Assignment
 Write an abstract of your thesis work (max. 250
words)
 Look for 2-3 works that applies fuzzy logic to your
thesis’ topic.
 Write a brief summary (max. 100 words/each)
 Look for 2-3 works that applies genetic algorithms to
your thesis’ topic .
 Look for 2-3 works that applies genetic fuzzy system to
your thesis’ topic .

Thank you very much
Any Question?

2015 Artificial Intelligence Techniques at Engineering Seminar - Chapter 2 - Part 4: Intelligent Transportation Systems

More Related Content

What's hot (20)

Similar to 2015 Artificial Intelligence Techniques at Engineering Seminar - Chapter 2 - Part 4: Intelligent Transportation Systems (20)

Recently uploaded (20)

2015 Artificial Intelligence Techniques at Engineering Seminar - Chapter 2 - Part 4: Intelligent Transportation Systems