Presentasjon lastes. Vennligst vent

Presentasjon lastes. Vennligst vent

Kunstig intelligens (MNFIT-272) - høst 2002. Leksjon 11 Emne: Subsymbolske metoder Nevrale nett - Biologisk basis - Perceptron - Backpropagation Evolusjonær.

Liknende presentasjoner


Presentasjon om: "Kunstig intelligens (MNFIT-272) - høst 2002. Leksjon 11 Emne: Subsymbolske metoder Nevrale nett - Biologisk basis - Perceptron - Backpropagation Evolusjonær."— Utskrift av presentasjonen:

1 Kunstig intelligens (MNFIT-272) - høst 2002. Leksjon 11 Emne: Subsymbolske metoder Nevrale nett - Biologisk basis - Perceptron - Backpropagation Evolusjonær komputasjon - Genetiske algoritmer - Genetisk programmering - Kunstig liv

2 har teknologisk perspektiv har metoder STUDIE AV INTELLIGENTE SYSTEMER RELATERT TIL KOMPUTASJONELLE PROSESSER REALISERE DATASYSTEMER SOM KAN SIES Å OPPVISE INTELLIGENT ADFERD - DVS. ' SMARTERE ' SYSTEMER KUNNSKAPSBASERTE METODER ADFERDSORIENTERTE METODER SUBSYMBOLSKE METODER KOGNITIV PSYKOLOGI FILOSOFI bygger bl.a. på MATEMATIKK BIOLOGI KUNSTIG INTELLIGENS (AI) har vitenskapelig perspektiv er koblet via empirisk vitenskapelig metode INFORMATIKK er delfelt av (Fra første forelesning)

3 SUBSYMBOLSKE METODER NEVRALE NETT Distribuert representasjon i et nettverk av noder koblet sammen via lenker (konneksjonisme). Ingen eksplisitt, direkte representasjon av begreper og sammenhenger. GENETISKE ALGORITMER Representasjon i form av bit-strenger som maniplueres vha. genetiske operatorer. Bit-strengene kan representere begreper, men ikke nødvendigvis.

4 Nevrale nett (NN) Alternativt AI-paradigme til kunnskaps- baserte systemer Distribuert - ikke-eksplisitt - representasjon er det mest typiske, men NN trenger ikke være det (lokalistiske nettverk) Kunnskapsrepresentasjons-hypotesen og fysisk-symbol-system hypotesen gjelder ikke for rene NN systemer Relaterte begreper: - Konneksjonsisme - Parallell distribuert prosessering (PDP)

5 Biologisk bakgrunn (løs): Et nettverk av nerveceller Cell body

6 NEVRALE NETT - struktur og læringsprinsipp Et nettverk av noder delvis koblet i sammen via lenker. Lenkene som går inn til en node kan medføre at noden aktiveres, som igjen medfører at lenken ut fra noden gis en verdi. En node i nettet: Aktiveringsverdien for en node (o) bestemmes av en aktiveringsfunksjon (f) der styrken på inn-lenkene (x..) samt av vektingen av hver av dem (w..) er input-parametre. Nevrale nett trenes opp ved at inndata gis inn, utdata registreres og avvikende resultatet tilbakeføres til nettet slik at vektene på lenkene blir justert ”i riktig retning” x1 x2 x4 x3 w1 w2 w3 w4 o f

7 Perceptron - enkelt-lags nett - binære input- og aktiverings-verider (-1 +1) - vekter er reelle tall - aktiveringsfunksjonen er er en trappe-terskelfunksjon - lærings-regel for å oppdatere vektene: w i = w i + c(d - o)x i w i = w i + c(d - sign(sum i x i w i ))x i - konvergerer hviss input-verdiene danner lineært separerbare klasser

8 Flerlags nett - Introduksjon av et skjult lag gjør at begrensningen om lineær separerbarhet kan omgås - input- og ut-verdier ikke kun binære - aktiveringsfunksjonen mer generell enn for Perceptroner - Backpropagation læringsmetode - en generalisering av regelen for Perceptroner -> Delta-regelen for læring - konvergens kan ikke garanteres generelt men for praktiske tilfeller vises

9 Egnede anvendelser av NN -Problemer der det er vanskelig å identifisere og beskrive kunnskap på eksplisitt form. -F.eks.: tolking av bilder, tolking og syntese av tale. Klassisk eksempel: NETTALK - System for kunstig tale (tekst -> tale) - Basert på Backpropagation - Oppsett lignende som CBRTALK (tidligere leksjon, se neste to transps) - Bestemmer fonem for bokstav midt i 7-bokstavs vindu, der fonemet avhenger av de 3 bokstaver på hver side - Nettkonfigurasjon: - 29 input enheter (26 bokstaver + ”.” ”,” ” ”) - 26 ut-enheter (21 lyder + 5 trykk og betoninger) - 3-lags nett med 80 skjulte noder

10 (Fra CBR tema - Memory-Based Reasoning)

11 (Fra CBR tema) MBR - (here: = case base)

12 Genetiske og Adferdsbaserte metoder -> Evolusjonære metoder Fra kognitive til biologiske - og etologiske - modeller av intelligent adferd. Intelligent adferd forårsakes ikke av modeller som er representert i maskinen, den oppstår i vekselvirkning med omgivelsene. Basis for det nye fagfeltet 'Artificial Life'.

13 Darwinian Evolution Genotypes Phenotypes Morphogenesis Natural Selection Recombination & Mutation Ptypes Gtypes Reproduction Sex Genetic Physiological, Behavioral

14 Evolutionary Algorithms Bit Strings Parameters, Code, Neural Nets, Rules Translate Performance Test Recombination & Mutation P,C,N,R Bits Generate Semantic Syntactic R &M

15 Evolutionary Computation = Parallel Stochastic Search 1 2 3 4 5 6 Biased Roulette Wheel CrossoverMutation Next Generation Selection Translation & Performance Test Selection Biasing

16 Types of Evolutionary Algorithms Genetic Algorithms (Holland, 1975) Representation: Bit Strings => Integer or real feature vectors Syntactic crossover (main) & mutation (secondary) Evolutionary Strategies (Recehenberg, 1972; Schwefel, 1995) Representation: Real-valued feature vectors Semantic mutation (main) & crossover (secondary) Evolutionary Programs (Fogel, Owens & Walsh, 1966; Fogel, 1995) Representation: Real-valued feature vectors or Finite State Machines Semantic mutation (only) View each individual as a whole species, hence no crossover Genetic Programs (Koza, 1992) Representation: Computer programs (typically in LISP) Syntactic crossover (main) & mutation (secondary)

17 Evolutionary Computation Requirements Domain that supports quantitative fitness assignment Fitness function that accurately evaluates performance Representation for solutions that tolerates mutation & crossover 011101001011011101001011 000111110010000111110010 7 4 11 1 15 2 P1 P2 P3 011101010010011101010010 000111101011000111101011 7 5 2 1 14 11 X Classic Genetic Algorithm

18 Travelling Salesman Problem (TSP) Given: N cities & matrix of distances between them. Find: Shortest cyclic tour that visits all cities. NP-Hard: Exponential to both find solutions & to verify solutions Heuristic Methods: Find optimal solutions when N< 1000. Genetic Algorithm: Find good solutions for any N Applications: Network building, Delivery routing, Sequence scheduling...

19 Applying GAs to TSP Fitness Function: 1/tour-length or optimal-tour-length/tour-length Chromosome: Direct Representation: List of cities (standard approach) 7 3 25 31 12 1 5 14…. Indirect Representation: List of next city to pull from ordered list and insert into the solution sequence 1 1 2 3 5…. => 1 2 4 6 9… Crossover Standard Bit or Integer Cross: Only works for indirect representations Location Preserving: Children inherit, as much as possible, cities in same gene location as parents Edge Preserving: Children inherit, as much as possible, city-city edges from parents (actual edge locations in the chromosome may vary from parent to kid) *Crossover is the key element to TSP GA’s - and where most research is done.

20 Representational Issues for GA-based TSP Standard Crossover & Mutation are purely syntactic => pay no attention to semantics (i.e. The meanings of the bits or integers that they manipulate). Direct representations often embody constraints that simple crossover & mutation cannot enforce. Indirect representations usually involve fewer constraints, so simple crossover & mutation are often sufficient. Compare to Process Scheduling (Kidwell), where constraints (i.e. All alleles between 0 and N) were easy to enforce. 4 2 3 5 4 1 6 3 2 1 5 66 3 2 5 4 1 4 2 3 1 5 6 2 x 6; 0 x 4 2 x 4; 0 x 6 X

21 Inheritance For GA’s to make progress, they must pass on many of the good features from generation G to generation G+1. Hence, when parents crossover, the good features of each should be preserved in at least one of the children. Biology: Heritability = degree to which children resemble their parents (X kid-avg - X pop-avg ) = heritability*(X parent-avg -X pop-avg ) Location Preservation (1985 - 1988) With PMX & Subtour exchanges, kids inherit many cities in the same position as in one of the parents. Results not that promising => GA viewed as inappropriate for TSP. Edge Preservation (1989 - present) Edges are the key contributors to TSP costs (fitness), so what we really need to preserve are city pairs (i.e. Edges) of TSP tours. GA’s with edge focus perform much better, near optimal => GA useful for TSP!

22 Using Evolutionary Algorithms When Large, rough search spaces Satisficing or Optimization problems Entire solutions are easily generated and tested Exhaustive search methods are too slow Heuristic search methods cannot find good solutions (e.g. Stuck at local max) How Determine EA-amenable representation of solutions Define fitness function Define selection function = roulette-wheel biasing function (f: fitness -> area) Set key EA parameters: population size, mutation rate, crossover rate, # generations, etc. * EA’s are easy to write, and there’s lots of freeware! * Specific problems often require specific representations & genetic operators

23 Application Areas for Evolutionary Algorithms Optimization: Controllers, Job Schedules, Networks(TSP) Electronics: Circuit Design (GP) Finance: Stock time-series analysis & prediction Economics: Emergence of Markets, Pricing & Purchasing Strategies Sociology: cooperation, communication, ANTS! Computer Science –Machine Learning: Classification, Prediction… –Algorithm design: Sorting networks Biology –Immunology: natural & virtual (computer immune system) –Ecology: arms races, coevolution –Population genetics: roles of mutation, crossover & inversion –Evolution & Learning: Baldwin Effect, Lamarckism…

24 Artificial Life Biology = study of carbon-based life –“life as we know it” Alife = study of the dynamics of living systems, regardless of substrate. –“life as it could be” –Substrates: abstract chemistries, logical networks, cellular automata, abstract ecosystems, emulated computers..

25 Emergence & Self-Regulation The signal feature of life is not the carbon-based substrate...(but)...that the local dynamics of a set of interacting entities (e.g. molecules, cells, etc.) supports an emergent set of global dynamical structures which stabilize themselves by setting the boundary conditions within which the local dynamics operates (Charles Taylor, biologist, UCLA) Global Structure Constraints Agents Emergence

26 Properties of Alife Systems Synthetic: Bottom-up, multiple interacting agents Self-regulating: No global/centralized control. Self-Organizing: Global structure is emergent. Adaptive: Learning and/or evolving Complex: On the edge of chaos; dissipative

27 Relevant Domains for Alife Individual organisms: physiology, behavior, ontogeny Social insect colonies (Ants!!): emergence of “super- organism” Populations of organisms: evolution Ecosystems: emergence & evolution MicroEconomics: populations of buyers & sellers Sociology: emergence & evolution of societies Traffic: emergence of flow/jam patterns Robotics: Emergence of Intelligent behaviors w/o global control Abstract computational models: logic, chemistry, physics

28 Why Study Alife? Understanding Emergent Phenomena –Synthetic approaches -vs- analytic reductionism –Chaos, complexity, self-organization Biological Research –Test effects of local behaviors upon populations –Genetic Engineering Computer Science –Nanotechnology, Animation –AI: Intelligence Architectures, Evolutionary Computation Educational Toolkits: –Social systems (SimCity) –Ecosystems (SimLife, SimEarth), Economical systems

29 What’s a living system? (Capra, “The Web of Life” (1996) Pattern: Autopoietic Network –Self-bounded –Self-generating –Self-perpetuating Structure: Dissipative –Far From Equilibrium –Edge of Chaos –Self-Organized Criticality Process: Cognition –Embodiment of pattern within structure –Life & Cognition are inseparable

30 Cellular Automata (ex.) Update rule: If exactly 2 dark neighbors, change to dark; else change to light. Step N Step N+1


Laste ned ppt "Kunstig intelligens (MNFIT-272) - høst 2002. Leksjon 11 Emne: Subsymbolske metoder Nevrale nett - Biologisk basis - Perceptron - Backpropagation Evolusjonær."

Liknende presentasjoner


Annonser fra Google