B.Y. Title: In telligen t Agen ts ✫ ✬ AIMA: Chapter 2 Choueiry In tro du tion to Arti� ial In telligen e CSCE 476-876, Spring 2016 URL: www. se.unl.edu/~ ho ue iry /S1 6- 476 -87 6 1 Berthe Y. Choueiry (Sh u-w e-ri) (402)472-5444 Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. In telligen t Agen ts ✫ ✬ Choueiry 1. Agen ts and en vironmen ts 2. Rationalit y 3. PEAS Sp e ifying the task en vironmen t: P erforman e measure, En vironmen t, A tuators, Sensors 2 4. T yp es of en vironmen ts 5. T yp es of In telligen t Agen ts Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
Agen t p er eiv es its en vironmen t through sensors An ything that B.Y. ✫ ✬ a ts up on its en vironmen t through a tuators Choueiry Agen ts in lude: Humans, rob ots, soft w are, et . Sensors? A tuators? 3 The agen t fun tion maps from p er ept sequen es to a tions: The agen t program runs on the ph ysi al ar hite ture to pro du e f Instru tor's f : P ∗ → A Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. ✫ ✬ V a uum- leaner w orld Choueiry A B P er epts: lo ations and on ten ts, e.g., [ A, dirty ] 4 A tions: Left , Right , Suck , NoOp Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
A V a uum- leaner Agen t P er ept sequen e A tion B.Y. Righ t ✫ ✬ Choueiry Su k Left Su k [ A, Clean ] , [ A, Clean ] Righ t [ A, Dirty ] . . . [ B, Clean ] , [ A, Clean ] , [ A, Clean ] Righ t [ B, Dirty ] . . [ A, Clean ] 5 . F un tion Re�ex-V a uum-Agen t ( [ location, status ]] ) returns an a tion if status = Dirty then return Suck [ A, Clean ] else if location = A then return Right else if location = B then return Left Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
Goal of AI B.Y. ✫ ✬ Build rational agen ts. Choueiry Rational = ? What is �rational� dep ends on: 1. P erforman e measures (ho w, when) 2. The agen ts' prior kno wledge of the en vironmen t 3. The a tions the agen t an p erform 6 4. P er ept sequen e to date (history): ev erything agen t has p er eiv ed so far Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. ✫ ✬ P erforman e meaure Choueiry Fixed p erforman e measure ev aluates the en vironmen t sequen e one p oin t p er square leaned up in time t p oin t p er lean square p er time step, min us one p er mo v e? 7 p enalize for > k dirt y squares? • • Instru tor's • Jan uary notes 25, 2016 #4 ✪ ✩
Rationalit y A rational agen t ho oses whi hev er a tion maximizes the B.Y. ✫ ✬ exp e ted v alue of the p erforman e measure giv en the p er ept Choueiry sequen e to date Rational � = omnis ien t, lairv o y an t Rationalit y maximizes exp e ted p erforman e P erfe tion maximizes a tual p erforman e Rational = exploration, learning, autonom y 8 After a su� ien t exp erien e of its en vironmen t, b eha vior of a rational agen ts b e omes e�e tiv ely indep enden t of prior kno wledge. ⇒ Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
PEAS B.Y. ✫ ✬ Choueiry T o design a rational agen t, w e m ust sp e ify the task en vironmen t P erforman e measure? En vironmen t? A tuators? Sensors? 9 Consider, e.g., the task of designing an automated taxi.. Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. PEAS : Automated taxi ✫ ✬ Choueiry P erforman e measure: safet y , destination, pro�ts, legalit y , omfort, . . . En vironmen t: US urban streets, freew a ys, tra� , p edestrians, stra y animals, w eather, . . . A tuators: steering, a elerator, brak e, horn, sp eak er/displa y , . . . 10 Sensors: video, a elerometers, gauges, engine sensors, k eyb oard, GPS, . . . Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. En vironmen t (1) ✫ ✬ Choueiry 1. F ully Observ able vs. P artially Observ able 2. Deterministi vs. sto hasti 3. Episo di vs. sequen tial 4. Stati vs. dynami 11 5. Dis rete vs. on tin uous 6. Single agen t vs. m ultiagen t Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
En vironmen t (2) B.Y. ✫ ✬ Choueiry F ully/P artially Observ able: sensors an dete t all asp e ts of the w orld E�e tiv ely fully observ able: relev an t asp e ts Deterministi vs. sto hasti : from the agen t's view p oin t Next state determined b y urren t state and agen ts' a tions P artially observ able + deterministi app ears sto hasti 12 Episo di vs. sequen tial: Agen t's exp erien e divided in to atomi episo des; subsequen t episo des do not dep end on a tions in previous episo des Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
En vironmen t (3) B.Y. ✫ ✬ Choueiry Stati vs. dynami : Dynami : En vironmen t hanges while agen t is delib erating Semidynami : en vironmen t stati , p erforman e s ores dynami Dis rete vs. on tin uous: Finite n um b er of pre epts, a tions Single agen t vs. m ultiagen t: B 's b eha vior maximizes a 13 p erforman e measure whose v alue dep ends on A 's b eha vior. Co op erativ e, omp etitiv e, omm uni ation. Chess? T axi driving? hardest ase? Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
En vironmen t (4) Hardest ase: patially observ able, sto hasti , sequen tial, dynami , B.Y. ✫ ✬ on tin uous, and m ultiagen t Choueiry Solitaire Ba kgammon In ternet shopping T axi Observ able Deterministi Episo di Stati Dis rete 14 Single-agen t Answ ers dep end on ho w y ou de�ne/in terpret the ase Episo di : hess tournamen t Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
En vironmen t t yp es B.Y. Solitaire Ba kgammon In ternet shopping T axi ✫ ✬ Choueiry Observ able Y es Y es No No Deterministi Y es No P artly No Episo di No No No No Stati Y es Semi Semi No Dis rete Y es Y es Y es No Single-agen t Y es No Y es No (ex ept au tions) 15 The en vironmen t t yp e largely determines the agen t design The real w orld is (of ourse) partially observ able, sto hasti , sequen tial, dynami , on tin uous, m ulti-agen t Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
B.Y. T yp es of Agen ts ✫ ✬ Choueiry F our, in order of in reasing generalit y: 1. Simple re�ex agen ts 2. Simple re�ex agen ts with state 3. Goal-based agen ts 4. Utilit y-based agen ts 16 5. Learning agen ts All these an b e turned in to learning agen ts. Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
Simple re�ex agen ts Simple lo ok-up table, mapping p er epts to a tions, is out of question B.Y. ✫ ✬ (to o large, to o exp ensiv e to build) Choueiry Man y situations an b e summarized b y ondition-a tion rules (h umans: learned resp onses, innate re�exes) • • Agent Sensors 17 What the world is like now Environment Re tangles: agen t's in ternal state Ov als: ba kground information What action I Condition-action rules Implemen tation: easy; Appli abilit y: narro w should do now Actuators Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
Simple re�ex agen ts with state B.Y. Sensory information alone is not su� ien t ✫ ✬ Choueiry Need to k eep tra k of ho w the w orld ev olv es (ev olution: indep enden tly of agen t, or aused b y agen t's a tions) • • Sensors 18 State What the world How the world evolves is like now Environment What my actions do Ho w the w orld ev olv ed: mo del-based agen t What action I Condition-action rules should do now Instru tor's Agent Actuators Jan uary notes 25, 2016 #4 ✪ ✩
Goal-based agen ts B.Y. State & a tions don't tell where to go ✫ ✬ Choueiry Need goals to build sequen es of a tions (planning) • • Sensors State What the world How the world evolves is like now 19 Environment What it will be like What my actions do if I do action A Goal-based: uses the same rules for di�eren t goals What action I Goals should do now Re�ex: will need a omplete set of rules for ea h goal Agent Actuators Instru tor's Jan uary notes 25, 2016 #4 ✪ ✩
Utilit y-based agen ts B.Y. Sev eral a tion sequen es to a hiev e some goal (binary pro ess) ✫ ✬ Choueiry Need to sele t among a tions & sequen es. Preferen es. Utilit y: State → real n um b er (express degree of satisfa tion, sp e ify trade-o�s b et w een on�i ting goal) • • • 20 � Sensors State What the world How the world evolves is like now Environment What it will be like What my actions do if I do action A How happy I will be Utility in such a state Instru tor's What action I Jan should do now uary Agent Actuators notes 25, 2016 #4 ✪ ✩
Recommend
More recommend
