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

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

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 Inputs Encoding � Concluding Remarks Notes on XCSs Notes on NXCSs MASSP NETTAB - July 19-21, 2002 2

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 3

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 4

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 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 � 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 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 � 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 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 � 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 8

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 9

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) α α α α α α α α α α − − β β β β β β β β β β − − − − β β β β β β β β β β c c A B C D A K L H I I B L M S R D F D S A MASSP NETTAB - July 19-21, 2002 10

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 11

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 overf it t ing ) MASSP NETTAB - July 19-21, 2002 12

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 e.g., t o classif y a pixel in a digit al image, a N 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) e.g., one or more f ilt ers can be applied t o a N given window of pixels. The result s summarize t he relevant f eat ures of t he window MASSP NETTAB - July 19-21, 2002 13

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 ? Cont ext ident if icat ion can be successf ully N exploit ed t o split t he input domain Regions − in t he case of secondary st ruct ures N predict ion − are input subsequences t hat show similar charact erist ics The “similarit y” crit eria act as cont ext N select ors MASSP NETTAB - July 19-21, 2002 14

The Pr oposed Solu ti on The Pr oposed Solu ti on The Pr oposed Solu ti on The Pr oposed Solu ti on ( Con ceptu a l Level) ( Con ceptu a l Level) ( Con ceptu a l Level) ( Con ceptu a l Level) Syntactic sugar? No, thanks. MASSP NETTAB - July 19-21, 2002 15

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 16

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 17

Zoom Ou t … Zoom Ou t … Population Ω of guarded experts output input Environment MASSP NETTAB - July 19-21, 2002 18

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 19

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 20

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 ( ) ( ) ( ) ( ) Γ = ◊ ⊥ x g x w x h x if then else where { } ( ) = ∈ = I g x I g x t r ue N g = boolean guard ⊆ ⊆ N h = t ot al or part ial f unct ion I I I g h N w = weight ing f unct ion 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 I I Guarded Expert x g h w enable (classifier / predictor) (weight) (guard) MASSP NETTAB - July 19-21, 2002 22

Ma cr o Ma cr o- - Ar chi tectu r e Ar chi tectu r e � Guarded experts can be arranged into a population … { } Ω = Γ Γ = = , w , i 1 , 2 ,..., g , h n i i i i i � Domain of a population of guarded experts { } ( ) ( ) ( ) Ω = Γ = ∈ ∃ = = U x I i 1 , 2 ,..., n g x t rue D D s.t. i i Γ ∈ Ω i MASSP NETTAB - July 19-21, 2002 23

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 24

The Pr oposed Solu ti on The Pr oposed Solu ti on 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) ( Ar chi tectu r a l Level) ( Ar chi tectu r a l Level) Experimenting multiple experts technology … MASSP NETTAB - July 19-21, 2002 25

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 26

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 = < h 1 , h 2 , … > MASSP NETTAB - July 19-21, 2002 27