Professional Documents
Culture Documents
by
Piero P. Bonissone & Kareem Aggour
Agenda
! Tuning Results
! Analysis & Conclusion
2
Motivation
! Driver reaction time often delays
identification of potentially hazardous
situations
3
Objective
Whew!
4
Related Work on FC for Braking
V = Vtracked (t ) Vtracking (t )
! Surface conditions (i.e., friction coefficient
between the vehicle tires and the road)
8
Simulator-FC-GA Architecture
Final Separation Distance
CARS SIMULATION
Difference Percentage of
in velocity brake pressure
KB
Scaling Factors
Rule Set
9
Tuning FC Knowledge Base with GA
! A Mamdani- type FLC approximates a relationship
between a state X and an output Y by using a KB
and a reasoning mechanism (interpreter)
implemented by the generalized modus-ponens.
! The Knowledge Base (KB) is defined by:
X1 X2
Very Low Low Medium High Very High
Very Low PH PH PM PL ZE
Changing Low PH PM PL ZE NL
a Term in Medium PM PL ZE NL NM
High PL ZE NL NM NH
X1 Very High ZE NL NM NH NH
X1 X2
Very Low Low Medium High Very High
Very Low PH PH PM PL ZE
Changing Low PH PM PL ZE NL
a Rule Medium PM PL ZE NL NM
High PL ZE NL NM NH
Very High ZE NL NM NH NH
FC Sensitivity to Parameter
and Structure Changes
X1 X2
Very Low Low Medium High Very High We use Non-
Very Low PH PH PM PL ZE
Changing Low PH PM PL ZE NL normalized
a SF Medium PM PL ZE NL NM
Term Sets, e,g,
High PL ZE NL NM NH
Very High ZE NL NM NH NH
the SF are
X1 X2
incorporated
Changing Very Low Low Medium High Very High
into the
a Term Very Low PH PH PM PL ZE
in a TS
Low
Medium
PH
PM
PM
PL
PL
ZE
ZE
NL
NL
NM
Term Sets.
High PL ZE NL NM NH
(e.g. X1) Very High ZE NL NM NH NH
So we will
X1
Very Low Low
X2
Medium High Very High
Focus our
Changing Very Low PH PH PM PL ZE tuning to the
a Rule Low PH PM PL ZE NL
in RS
Medium PM PL ZE NL NM Term Sets
High PL ZE NL NM NH
Very High ZE NL NM NH NH
Simulator-FC-GA Architecture
CARS
FC1 & FC2 SIMULATION
Decoder
Termsets
Fitness
Function
Final Separation Distance
Percentage
Percentage
Difference of brake
of brake
in velocity pressure
FC1(ideal FC2
pressure
Separation (actual
surface)
Distance surface)
Vo + V f
d= t V f = Vo + at
2
1
momentum= mV 2
Workfriction = mgd
2
Work net = Momentum
1 1
Work friction = mV f
2
mV o
2
2 2
1
mgd = mV o
2
2
14
Simulator Design Equations
Vo
2 2
V
d stopping = o
a =
2g 2d
a maximum = g
V ( t + 1 ) = V ( t ) + ( a maximum % Braking ) t
15
Fuzzy Controller(s) Design
! Designed system containing two fuzzy controllers
(FC) connected in series
Difference in
velocity % Pressure
(Ideal surface)
Separation Fuzzy Controller 1
% Pressure
distance
Fuzzy Controller 2 (Actual surface)
s
17
Rule Sets FC 2
Coefficient of Friction Term Set Term Sets
% Brake Pressure T. Set
State Variables:
M%P F%P VH%P H%P M%P M%P % PIdeal
18
Genetic Algorithm
Tuning Term Sets
! Chromosome Design:
! Parametric representation of the term sets:
! Slopes intersection at 0.5, and core (1 side)
(Generic
Term Set)
0
! Guarantee that for each point in the universe of discourse the sum
of its associated membership values equal one
! Reduces degrees of freedom (by )
! Preserved a good property of the term set, which usually provides
good interpolation behavior
19
GA Parameters
! Encoding = real-valued ! Elitist strategy = YES
! Population size = 200 ! Crossover type = single cut
! Probability Mutation = 0.5% ! Maximum number generations
! Probability Crossover = 30% = 2,000
! Fitness Scenarios = 15 ! Selection type = Rank
Fitness Function:
Final <= (-10, (-7.5, (-5, (-2, (0, (1, (3.5, (5, (7, (10, 12
(2.5, 3.5]
Separation -10 -7.5] -5] -2] 0] 1] 2.5] 5] 7] 10] 12] <=
Reward -25 -20 -15 -10 -5 5 8 15 10 5 2 -2 -5
20
Tuned Term Sets
1.2 1.2
VL L M H VH VL L M H VH
0.8 0.8
Membership
Membership
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 10 20 30 40 50 60 70 0 20 40 60 80 100
VL L M H VH Z VL L M H VH F
0.8
0.8
Membership
Membership
0.6
0.6
0.4
0.4
0.2
0.2
21
0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Brake Pressure (%)
Coefficient of Friction (mu)
Tuned Term Sets (Friction Coefficients)
1.2
Coefficient of Friction
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Ice
Coefficient of Friction (mu) Dry Concrete
Snow Dry Asphalt
Wet Asphalt Wet Concrete 22
Analysis & Conclusion
! Performance:
! Tested against hundreds of random scenarios with wide
range of initial distances, velocity and friction coefficients
! Consistently stopped within 3m ( 0.5m) of the tracked vehicle
(when physically possible)
! Occasional accidents were possible only when tracked vehicle
defied laws of physics due to unrealistic random initial conditions
! Two controllers in series worked well
! Reduced size of rule sets by ~ 50%
! # Rules in cascading controllers: 5X5 + 7x5 = 60
! # Rules in a 3-input controller: 5X5X5 = 125
! The GA tuning of the Term Sets was very successful and
considerably improved performance of the manually
designed FCs
23
Comments for future Work
! L could be a function of Delta V
! We could add a second goal of tracking
vehicle at safe distance
24