An Agent Architecture An Agent Architecture An Agent Architecture - - PowerPoint PPT Presentation

An Agent Architecture An Agent Architecture for Predicting Protein Secondary for Predicting Protein Secondary Structures Structures An Agent Architecture An Agent Architecture for Predicting Protein Secondary for Predicting Protein Secondary Structures Structures

• G. Ar m a n o, ( *) L. Mi l a n esi , ( ^ ) a n d A. Or r o ( *)

(*) DIEE - University of Cagliari, Cagliari, Italy email: {armano,orro}@diee.unica.it (^ ) ITB – CNR, Milano, Italy email: milanesi@ itba.mi.cnr.it

MASSP NETTAB - July 19-21, 2002 2

Ou tli n e of the Ta lk Ou tli n e of the Ta lk

Introduction Focusing on the Problem … The Proposed Solution (Conceptual Level) The Proposed Solution (Architectural Level) The Proposed Solution (Design Level) Experimental Results Concluding Remarks

Inputs Encoding Notes on XCSs Notes on NXCSs

MASSP NETTAB - July 19-21, 2002 3

I n tr od u cti on I n tr od u cti on I n tr od u cti on I n tr od u cti on

…

MASSP NETTAB - July 19-21, 2002 4

Why Pr ed i cti n g Secon d a r y Str u ctu r es ? Why Pr ed i cti n g Secon d a r y Str u ctu r es ?

Finding the actual labeling through existing techniques may become too expensive if performed on a large scale Predicting the actual labeling is less expensive …

MASSP NETTAB - July 19-21, 2002 5

Exi sti n g Method s for Pr ed i cti n g Exi sti n g Method s for Pr ed i cti n g Secon d a r y Str u ctu r es Secon d a r y Str u ctu r es

Purely syntactic methods

N Based on t he analysis of t he primary st ruct ure

perf ormed using grammar-based and / or machine learning approaches

Comparative Modeling Fold Recognition Ab-initio Methods

MASSP NETTAB - July 19-21, 2002 6

Exi sti n g Method s for Pr ed i cti n g Exi sti n g Method s for Pr ed i cti n g Secon d a r y Str u ctu r es Secon d a r y Str u ctu r es

Purely syntactic methods Comparative Modeling

N Based on t he similarit y bet ween t est sequences and

t he ones available in st ruct ural dat abases

Fold Recognition Ab-initio Methods

MASSP NETTAB - July 19-21, 2002 7

Exi sti n g Method s for Pr ed i cti n g Exi sti n g Method s for Pr ed i cti n g Secon d a r y Str u ctu r es Secon d a r y Str u ctu r es

Purely syntactic methods Comparative Modeling Fold Recognition

N Based on st ruct ural t emplat es whose mat ching

sequences have a known spat ial f olding

Ab-initio Methods

MASSP NETTAB - July 19-21, 2002 8

Exi sti n g Method s for Pr ed i cti n g Exi sti n g Method s for Pr ed i cti n g Secon d a r y Str u ctu r es Secon d a r y Str u ctu r es

Purely syntactic methods Comparative Modeling Fold Recognition Ab-initio Methods

N Use a lat t ice model t o predict t he st ruct ure by

minimizing an energy f unct ion

MASSP NETTAB - July 19-21, 2002 9

Focu si n g on the Pr oblem … Focu si n g on the Pr oblem … Focu si n g on the Pr oblem … Focu si n g on the Pr oblem …

Let’s get to the point …

MASSP NETTAB - July 19-21, 2002 10

Focu si n g on the Pr oblem … Focu si n g on the Pr oblem …

Given an amino acidic sequence, predict its secondary structure (α-helix, β-sheet, or coil)

A B C D A K L H I I B L M S R D F D S A α α α α α α α α α α − − β β β β β β β β β β − − c c − − β β β β β β β β β β

MASSP NETTAB - July 19-21, 2002 11

Usi n g a Globa l Mod el ? Usi n g a Globa l Mod el ?

Global models …

N Of t en rely on a “st at e-based” approach (e.g.,

HMMs, Recurrent ANNs)

N Must be t rained on large input sequences, t o

(hopef ully) be able t o ident if y t he underlying syst em

N Lack of generalizat ion abilit y (in t erms of

underf it t ing)

MASSP NETTAB - July 19-21, 2002 12

Usi n g Loca l Mod els ? Usi n g Loca l Mod els ?

Local models …

N Do not require a “st at e-based” approach (t hey

can be “cont ext -based”)

N Do not require t o be t rained on large input

sequences

N Lack of generalizat ion abilit y (in t erms of

• verf it t ing )

MASSP NETTAB - July 19-21, 2002 13

Con text Con text-

• vs. Sta te
• vs. Sta te-
• Ba sed Appr oa ch

Ba sed Appr oa ch

Contexts usually apply to classification tasks

N

e.g., t o classif y a pixel in a digit al image, a limit ed window of surrounding pixels can be t aken int o account

Contexts may be summarized by suitable metrics (thus reducing the complexity of the learning task)

N

e.g., one or more f ilt ers can be applied t o a given window of pixels. The result s summarize t he relevant f eat ures of t he window

MASSP NETTAB - July 19-21, 2002 14

Con text Con text-

• Ba sed Cla ssi fi ca ti on

Ba sed Cla ssi fi ca ti on

Why adopting a “context-based” approach also for prediction tasks ?

N

Cont ext ident if icat ion can be successf ully exploit ed t o split t he input domain

N

Regions −in t he case of secondary st ruct ures predict ion− are input subsequences t hat show similar charact erist ics

N

The “similarit y” crit eria act as cont ext select ors

MASSP NETTAB - July 19-21, 2002 15

The Pr oposed Solu ti on The Pr oposed Solu ti on ( Con ceptu a l Level) ( Con ceptu a l Level) The Pr oposed Solu ti on The Pr oposed Solu ti on ( Con ceptu a l Level) ( Con ceptu a l Level)

Syntactic sugar? No, thanks.

MASSP NETTAB - July 19-21, 2002 16

Solu ti on ( Con ceptu a l Level) Solu ti on ( Con ceptu a l Level)

Using local models (context-based approach) Devising a population of experts Each expert participates to the prediction process only on a (usually small) subset of the input sequences

MASSP NETTAB - July 19-21, 2002 17

Un d er lyi n g Assu m pti on Un d er lyi n g Assu m pti on

Splitting the input space allows to make it easier the classification task, in a multiple- experts perspective

MASSP NETTAB - July 19-21, 2002 18

Zoom Ou t … Zoom Ou t …

Population Ω of guarded experts Environment input

• utput

MASSP NETTAB - July 19-21, 2002 19

Zoom I n … Zoom I n …

Micro-Architecture …

N Def ining guarded expert s

Macro-Architecture …

N Handling a populat ion of guarded expert s

MASSP NETTAB - July 19-21, 2002 20

Mi cr o Mi cr o-

• Ar chi tectu r e

Ar chi tectu r e

A guarded expert is a triple <g,h,w> where:

N h is a t ot al or part ial f unct ion t hat maps an input

space (I ) t o an out put space (O)

N g is a boolean f unct ion devot ed t o cont rol t he

act ivat ion of h (i.e., g is a “guard” t hat ident if ies a subset of input s f or which t he mapping exist s)

N w is a weight ing f unct ion, which ident if ies t he

st rengt h of t he expert

MASSP NETTAB - July 19-21, 2002 21

Mi cr o Mi cr o-

• Ar chi tectu r e

Ar chi tectu r e

• I I

I I

In symbols: Γ = < g,h,w>= guarded expert Γ : I g → O, D(Γ) = I g

where

N g = boolean guard N h = t ot al or part ial f unct ion N w = weight ing f unct ion

( ) ( ) ( ) ( )

⊥ ◊ = else then if x h x w x g x Γ

I I I

h g

⊆ ⊆

( )

{ }

t r ue x g I x I g = ∈ =

MASSP NETTAB - July 19-21, 2002 22

Mi cr o Mi cr o-

• Ar chi tectu r e

Ar chi tectu r e

• I I I

I I I

(classifier / predictor)

Guarded Expert

x enable

g h

(guard) (weight)

w

MASSP NETTAB - July 19-21, 2002 23

Ma cr o Ma cr o-

• Ar chi tectu r e

Ar chi tectu r e

Guarded experts can be arranged into a population … Domain of a population of guarded experts

{ }

n h , g

i i

,..., , i , w ,

i i i

2 1 = = = Γ Γ Ω

( )

( )

( )

{ }

t rue x g n ,..., , i I x

i i

i

= = ∃ ∈ = =

∈

s.t. 2 1

U

Ω Γ

Γ Ω D D

MASSP NETTAB - July 19-21, 2002 24

Ma cr o Ma cr o-

• Ar chi tectu r e

Ar chi tectu r e

• I I

I I

To handle a population of guarded experts several decision must be taken:

N Training st rat egy and t echnique ? N Region Split t ing Crit eria (boundaries and overlapping) ? N Expert s Select ion Mechanism (usually required) ? N Out put s Blending Mechanism (usually required) ? N Vot ing Policy (usually required) ?

MASSP NETTAB - July 19-21, 2002 25

The Pr oposed Solu ti on The Pr oposed Solu ti on ( Ar chi tectu r a l Level) ( Ar chi tectu r a l Level) The Pr oposed Solu ti on The Pr oposed Solu ti on ( Ar chi tectu r a l Level) ( Ar chi tectu r a l Level)

Experimenting multiple experts technology …

MASSP NETTAB - July 19-21, 2002 26

A Hybr i d Ar chi tectu r e for A Hybr i d Ar chi tectu r e for Pr ed i cti n g Secon d a r y Str u ctu r es Pr ed i cti n g Secon d a r y Str u ctu r es

Micro-Architecture …

N Devising a hybrid guarded expert using eXt ended

Classif ier Syst ems (XCSs) and Art if icial Neural Net works (ANNs)

Macro-Architecture …

N I mplement ing t he populat ion of expert s as a

societ y of agent s

N Using simple coordinat ion policies

XCS ≈ ≈ reinforcement learning + genetic algorithms

MASSP NETTAB - July 19-21, 2002 27

Mi cr o Mi cr o-

• Ar chi tectu r e: NXCS Exper ts

Ar chi tectu r e: NXCS Exper ts

A “Neural XCS” expert (NXCS expert for short) is a Guarded Expert where …

N g = an XCS-like classif ier (maps it s input s t o bool) N h = an ANN (suit ably cust omized) N w = expert ’s f it ness

In case of multiple outputs …

N h = <h1, h2, …

>

MASSP NETTAB - July 19-21, 2002 28

Ma cr o Ma cr o-

• Ar chi tectu r e:

Ar chi tectu r e: Ha n d li n g a Ha n d li n g a Popu la ti on of NXCS Exper ts Popu la ti on of NXCS Exper ts

Training strategy: batch Training technique: Darwinian selection (XCS guards) + backpropagation (ANNs) Region splitting: hard, with overlapping Experts’ Selection: match-set formation Outputs blending: fitness-weighted averaging Voting: plurality rule

MASSP NETTAB - July 19-21, 2002 29

Selecti on , Ou tpu ts Blen d i n g, a n d Selecti on , Ou tpu ts Blen d i n g, a n d Voti n g Voti n g

where …

N Ω denot es t he populat ion of expert s N x denot es t he current input N select () creat es t he mat ch-set N combine() blends out put s of all expert s t hat

belong t o t he mat ch-set

N choose() enf orces t he adopt ed vot ing policy

( ) ( )

( ) ( ) x

Ω select combine choose x O =

MASSP NETTAB - July 19-21, 2002 30

Selecti on , Ou tpu ts Blen d i n g, a n d Selecti on , Ou tpu ts Blen d i n g, a n d Voti n g Voti n g

• I I

I I

Selection (match-set formation)

N

Mat ch Set

Outputs blending (fitness-weighted average)

N

Overall Out put

N

k ∈ { α, β, c }

Voting (plurality rule)

N choose(O) →

Select ed Out put { }

k c , , k

• max

ar g * k

β α ∈

=

( )

{ }

M e ,..., e , e select

L x

= →

2 1

Ω

( )

∑ ∑

∈ ∈

⋅ =

M e e M e ek e k

f x h f

• (

)

( )

x O

• ,
• ,
• M

combine

c x

= →

β α

MASSP NETTAB - July 19-21, 2002 31

I n pu ts to a Gen eti c Gu a r d I n pu ts to a Gen eti c Gu a r d

A pattern of physico-chemical properties is generated, to be matched with a moving window of residuals

MASSP NETTAB - July 19-21, 2002 32

I n pu ts to a Gen eti c Gu a r d I n pu ts to a Gen eti c Gu a r d

A pattern of physico-chemical properties is generated, to be matched with a moving window of residuals

MASSP NETTAB - July 19-21, 2002 33

I n pu ts to a Gen eti c Gu a r d I n pu ts to a Gen eti c Gu a r d

A pattern of physico-chemical properties is generated, to be matched with a moving window of residuals Acidity

N Acid (+) N Basic (-) N Neutral (=)

Hydrophobicity

N Hydrophobic N Hydrophilic

Polarity

N Polar N Aromatic N Aliphatic

MASSP NETTAB - July 19-21, 2002 34

I / O of a Neu r a l Pr ed i ctor I / O of a Neu r a l Pr ed i ctor

α α β β c

central residue moving window 2+1+2 central residue

amino acidic sequence actually 10+1+10

MASSP NETTAB - July 19-21, 2002 35

I / O of a Neu r a l Pr ed i ctor I / O of a Neu r a l Pr ed i ctor

• I I

I I

Inputs

N A moving window of 21 = 10+1+10 residues

Outputs

N Three separat e out put s: α-helix, β-sheet , coil N Current ly, no rej ect ion opt ion

central residue

MASSP NETTAB - July 19-21, 2002 36

The Pr oposed Solu ti on The Pr oposed Solu ti on ( Desi gn Level) ( Desi gn Level) The Pr oposed Solu ti on The Pr oposed Solu ti on ( Desi gn Level) ( Desi gn Level)

Experimenting agent-oriented technology …

MASSP NETTAB - July 19-21, 2002 37

Agen t Agen t-

• Or i en ted I m plem en ta ti on of the

Or i en ted I m plem en ta ti on of the Pr oposed Ar chi tectu r e Pr oposed Ar chi tectu r e

Population Ω of NXCS experts Environment reward

• utput

Creation Manager Selector Rewarding Manager Combination Manager input

MASSP NETTAB - July 19-21, 2002 38

I m plem en ta ti on of Agen ts I m plem en ta ti on of Agen ts

Coordination agents:

N Select or N Combinat ion Manager N Rewarding Manager N Creat ion Manager

NXCS experts population

N A “societ y” of NXCS expert s

MASSP NETTAB - July 19-21, 2002 39

I m plem en ta ti on of Agen ts I m plem en ta ti on of Agen ts

Both coordination agents and NXCS experts are implemented as active objects Each active object embodies two interacting subsystems

N a communicat ion subsyst em (ent rust ed wit h I / O) N an engine (ent rust ed wit h operat ions execut ion)

MASSP NETTAB - July 19-21, 2002 40

I m plem en ta ti on of Agen ts : the I m plem en ta ti on of Agen ts : the Com m u n i ca ti on Su bsystem Com m u n i ca ti on Su bsystem

Equipped with an input queue, hosting incoming messages Output messages are not queued. Synchronous and asynchronous message passing

MASSP NETTAB - July 19-21, 2002 41

I m plem en ta ti on of Agen ts : the I m plem en ta ti on of Agen ts : the En gi n e En gi n e

Non-preemptive thread scheduler (writers are mutually exclusive, whereas readers are not) Reactive behavior is hand-coded

N The current t ask can be suspended and possibly

resumed af t er serving t he int errupt (no goal st acking, t hough!)

Proactive behavior is hand-coded

N The current goal is select ed among a predef ined

set of “behaviors”

MASSP NETTAB - July 19-21, 2002 42

En for ci n g Coor d i n a ti on Poli ci es En for ci n g Coor d i n a ti on Poli ci es

Selector

N Once inf ormed t hat t here is an input t o process,

af t er a preliminary check, aimed at f orming t he mat ch set , t he select or int eract s wit h t he plat f orm of NXCS expert s

N The mat ch set (t oget her wit h t he current input )

is f orwarded t o t he Combinat ion Manager

MASSP NETTAB - July 19-21, 2002 43

En for ci n g Coor d i n a ti on Poli ci es En for ci n g Coor d i n a ti on Poli ci es

Combination Manager

N Af t er receiving inf ormat ion about t he mat ch set

and t he current input , t he combinat ion manager produces t he f inal out put by int eract ing wit h all NXCS expert s t hat belong t o t he mat ch set

N I n t he special case of an empt y mat ch set , t he

combinat ion manager eit her inf orms t he creat ion manager t hat a new expert , able t o cover t he current input , must be creat ed (t raining st at us)

• r out put s a “void” predict ion (t est st at us)

MASSP NETTAB - July 19-21, 2002 44

En for ci n g Coor d i n a ti on Poli ci es En for ci n g Coor d i n a ti on Poli ci es

Rewarding Manager

N I nf orms all NXCS expert s belonging t o t he

mat ch set t hat t hey should updat e t heir f it ness, according t o t he reward obt ained by t he environment (t raining st at us)

MASSP NETTAB - July 19-21, 2002 45

En for ci n g Coor d i n a ti on Poli ci es En for ci n g Coor d i n a ti on Poli ci es

Creation Manager

N Handles expert s’ creat ion, being able t o perf orm

covering, crossover, and mut at ion operat ions

N I t is also responsible f or expert s delet ion, wit h a

probabilit y inversely proport ional t he f it ness of each expert

MASSP NETTAB - July 19-21, 2002 46

Exper i m en ta l Resu lts Exper i m en ta l Resu lts Exper i m en ta l Resu lts Exper i m en ta l Resu lts

MASSP NETTAB - July 19-21, 2002 47

Exper i m en ts Exper i m en ts

About 126 training sequences have been taken into account

N RS126 dat aset

About 396 test sequences have been taken into account

N CB396 dat aset

MASSP NETTAB - July 19-21, 2002 48

Metr i cs Ad opted to Assess the Pr ed i cti on Metr i cs Ad opted to Assess the Pr ed i cti on Ca pa bi li ty Ca pa bi li ty

Q3

N Percent of residuals correct ly predict ed vs. t he

• verall number of residuals

SOV (Segment Overlap)

N Measures, f or each conf ormat ional st at us, t he

• verlapping bet ween predict ed and correct element s

MASSP NETTAB - July 19-21, 2002 49

Com pa r a ti ve Resu lts Com pa r a ti ve Resu lts

RS1 2 6 protein set CB3 9 6 protein set Method Q3 SOV Q3 SOV PHD 73.5 73.5 71.9 75.3 DSC 71.1 71.6 68.4 72.0 PREDATOR 70.3 69.9 68.6 69.8 NNSSP 72.7 70.6 71.4 71.3 CONSENSUS 74.8 74.5 72.9 75.4 MASSP 71.4 68.9 69.1 70.5

MASSP NETTAB - July 19-21, 2002 50

Con clu d i n g Rem a r ks Con clu d i n g Rem a r ks Con clu d i n g Rem a r ks Con clu d i n g Rem a r ks

MASSP NETTAB - July 19-21, 2002 51

Con clu si on s Con clu si on s

To tackle the problem of predicting amino acidic secondary structures …

N We experiment ed t he mult iple expert s

t echnology

N Their implement at ion f ollowed t he guidelines of

agent -orient ed programming

Results are encouraging, although the focus was on experimenting new technologies rather than on improving other systems’ performances

MASSP NETTAB - July 19-21, 2002 52

Fu tu r e Wor k Fu tu r e Wor k

The next release of the system, able to implement a subset of FIPA ACL is currently under way The final architecture will be a society of heterogeneous agents headed at predicting the secondary structure of amino acidic sequences

MASSP NETTAB - July 19-21, 2002 53

Well… Well…

Questions ?

MASSP NETTAB - July 19-21, 2002 54

I n pu ts En cod i n g I n pu ts En cod i n g I n pu ts En cod i n g I n pu ts En cod i n g

MASSP NETTAB - July 19-21, 2002 55

I n pu ts En cod i n g I n pu ts En cod i n g

Amino acids are encoded according to the following physico-chemical properties

N Acidit y N Hydrophobicit y N Polarit y

MASSP NETTAB - July 19-21, 2002 56

I n pu ts En cod i n g I n pu ts En cod i n g

• I I

I I

Acidity

N Acid

(+) 001

N Basic

(-) 010

N Neutral

(=) 100

Hydrophobicity

N Hydrophobic

01

N Hydrophilic

10

Polarity

N Polar

001

N Aromatic

010

N Aliphatic

100

N Other

000

MASSP NETTAB - July 19-21, 2002 57

I n pu ts En cod i n g I n pu ts En cod i n g

• I I I

I I I

A 100 01 100 (a) R 010 10 000 (b) N 100 10 001 (c) D 001 10 000 (d) C 100 10 001 (c) Q 100 10 001 (c) E 001 10 000 (d) G 100 10 100 (e) H 010 10 000 (b) I 100 01 100 (a) L 100 01 100 (a) K 010 10 000 (b) M 100 01 001 (f) F 100 01 010 (g) P 100 01 100 (a) S 100 10 001 (c) T 100 10 001 (c) W 100 01 010 (g) Y 100 10 010 (h) V 100 01 100 (a) (a) / 5, (b) / 3, (c) / 5, (d) / 2, (e) / 1, (f) / 1, (g) / 2, (h) / 1

MASSP NETTAB - July 19-21, 2002 58

Notes on NXCS Notes on NXCS Notes on NXCS Notes on NXCS

MASSP NETTAB - July 19-21, 2002 59

NXCS– TYPI CAL TRAI NI NG LOOP

0. Start with an empty or existing population of experts. 1. Given an input x, build the match set M. 2. If M is empty, generate a new expert able to cover x. 3. Select an action a* according to a suitable policy (typically, a fitness-weighted majority / plurality rule). 4. Update p, ε ε, and f of each classifier in M. 5. When needed, generate a new pair of experts using genetic

• perators (crossover and mutation). Insert the pair of

predictors in the population. 6. When needed, delete a pair of experts from the population. 7. Go to Step 1.

MASSP NETTAB - July 19-21, 2002 60

Notes on XCSs Notes on XCSs Notes on XCSs Notes on XCSs

MASSP NETTAB - July 19-21, 2002 61

XCS Key Con cepts XCS Key Con cepts

An XCS is an evolutionary learning system consisting of the following components …

N perf ormance N reinf orcement N discovery

MASSP NETTAB - July 19-21, 2002 62

Ba si c XCS Ar chi tectu r e Ba si c XCS Ar chi tectu r e

evolut ionary behavior

ENVIRONMENT

input

• ut put

reward mat ching act ion select ion

Population

• f

XCS XCS classifiers Match Set Action Set

rewarding

MASSP NETTAB - July 19-21, 2002 63

XCS Cla ssi fi er s XCS Cla ssi fi er s -

• Ma i n Pa r a m eter s

Ma i n Pa r a m eter s

The most important parameters of an XCS classifier are:

N Predict ion ( p ) N Predict ion Error ( ε ) N Fit ness ( f ) N Relat ive Accuracy ( k’ ) N Accuracy ( k )

( )

1 1 + + = t t t

p , r , ε λ ε

ε

( )

1 1 + +

′ =

t t f t

k , r , f f λ

( )

* a

M t k t

, k k Κ λ

1 1 + +

′ = ′

( )

1 +

=

t k

k ε λ

( )

r , p p

t p t

λ =

+1

MASSP NETTAB - July 19-21, 2002 64

XCS Cla ssi fi er s XCS Cla ssi fi er s -

• Ma i n Pa r a m eter s

Ma i n Pa r a m eter s

( )

t t 1 t

p r p p − ⋅ + =

+

β

( )

t 1 t t 1 t

r p ε β ε ε − − ⋅ + =

+ +

( )

t t 1 t

f k f f − ′ ⋅ + =

+

β

∑

∈

= ′

* a

c c

k k k

M

Operation Formula Where … Prediction Update (single step) r is the actual reward

• btained by the system from

the environment. Prediction Error Update Fitness Update k’ is the relative accuracy of a classifier. Relative Accuracy The accuracy of a classifier is normalized over the action set corresponding to the selected action a*. Accuracy α = accuracy fall-off ε0 = accuracy threshold (note that k=α when ε = 2ε0).

      − ⋅ = ) ln( exp k ε ε ε α

MASSP NETTAB - July 19-21, 2002 65

XCS– TYPI CAL TRAI NI NG LOOP

0.

Start with an empty, or existing, population of classifiers. 1. Given an input x, build the match set M. 2. If M is empty, generate a new classifier able to cover x. 3. Select an action a* according to a suitable strategy (typically, a fitness weighted majority rule). 4. Update p, ε ε, and f of each classifier that supports a*. 5. When needed, generate a new pair of classifiers using standard genetic operators (crossover and mutation). Insert the pair of classifiers in the population. 6. When needed, a pair of classifiers is deleted. 7. Go to Step 1.