Transcript Artificial Intelligence - TAMU Computer Science Faculty Pages

Artificial Intelligence
What is AI?
•
•
•
•
•
Can machines “think”?
Can machines be truly autonomous?
Can machines program themselves?
Can machines learn?
Will they ever be “conscious”, and is that
necessary?
• media depictions of AI (science fiction)
– HAL in 2001: Space Odessey
– Spielberg’s AI
– Data on Star Trek Next Generation
– Terminator, Skynet...
• real AI has many practical applications
– credit evaluation, medical diagnosis
– guidance systems, surveillance
– manufacturing (robotics, logistics)
– information kiosks, computer-aided tutoring
– AI in video games (also: Deep Blue, chess)
– driverless vehicles, UAVs
– Mars rover, Hubble telescope
• AI has a long history, and draws on many
fields
– mathematics, computability, formal logic
– control theory
– optimization
– cognitive science
– linguistics
Perspectives on AI
• Philosophical
– What is the nature of intelligence?
• Psychological
– How do humans think?
• Engineering
– advanced methods for building complex
systems that solve hard real-world problems
Philosophical Perspective
• started with Greek philosophers (e.g. Aristotle)
– syllogisms
– natural categories
• 1700-1800s: development of logic, calculus
– Descartes, Liebnitz, Boole, Frege, Tarski, Russell
– what are concepts? existence, intention, causality...
– reductionist approaches to try to mechanize reasoning
• 1900s: development of computers
– input/output model
– is intelligence a “computable function”?
– Turing, von Neumann, Gödel
• Does “intelligence” require a physical brain?
– Programmed devices will probably never have
“free will”
• Or is it sufficient to produce intelligent
behavior, regardless of how it works?
– operational definition
• The Turing Test
– first proposed in 1950 by Alan Turing
– a panel of human judges asks questions through
a teletype interface (restricted to topic areas)
– a program is intelligent if it can fool the judges
and look indistinguishable from humans
– chatterbots
– the Loebner Prize: annual competition at MIT
Psychological Perspective
• AI is about understanding and modeling
human intelligence
• Cognitive Science movement (ca. 1950s)
– replace stimulus/response model
– internal representations
– mind viewed as “information processor”
(sensory perceptionsconceptsactions)
• Are humans a good model of intelligence?
– strengths:
• Are humans a good model of intelligence?
– strengths:
•
•
•
•
interpretation, dealing with ambiguity, nuance
judgement (even for ill-defined situations)
insight, creativity
adaptiveness
• Are humans a good model of intelligence?
– strengths:
•
•
•
•
interpretation, dealing with ambiguity, nuance
judgement (even for ill-defined situations)
insight, creativity
adaptiveness
– weaknesses:
• Are humans a good model of intelligence?
– strengths:
•
•
•
•
interpretation, dealing with ambiguity, nuance
judgement (even for ill-defined situations)
insight, creativity
adaptiveness
– weaknesses:
•
•
•
•
•
slow
error-prone
limited memory
subject to biases
influenced by emotions
Optimization
• AI draws upon optimization methods
– remember NP-hard problems?
• there is (probably) no efficient algorithm that solves them in
polynomial time
– but we can have approximation algorithms
• run in polynomial time, but don’t guarantee optimal solution
– classic techniques: linear programming, gradient descent
• Many problems in AI are NP-hard (or worse)
• AI gives us techniques for solving them
–
–
–
–
heuristic search
use of expertise encoded in knowledge bases
AI relies heavily on greedy algorithms, e.g. for scheduling
custom algorithms for search (e.g. constraint
satisfaction), planning (e.g. POP, GraphPlan), learning
(e.g. rule generation), decision making (MDPs)
Planning
• Autonomy – we want computers to figure out
how to achieve goals on their own
– Given a goal G
– and a library of possible actions Ak
– find a sequence of actions A1..An
– that changes the state of the worlds to achieve G
current state of world
desired state of world
Planning
• Autonomy – we want computers to figure out
how to achieve goals on their own
– Given a goal G
– and a library of possible actions Ak
– find a sequence of actions A1..An
– that changes the state of the worlds to achieve G
pickup(A)
puton(A,table)
pickup(C)
puton(C,A)
pickup(B)
current state of world
pickup(B,C)
desired state of world
• Examples:
–
–
–
–
Blocks World – stack blocks in a desired way
traveling from College Station to Statue of Liberty
rescuing a victim in a collapsed building
cooking a meal
• The challenges of planning are:
– for each task, must invoke sub-tasks to ensure
pre-conditions are satisfied
• in order to nail 2 pieces of wood together, I have to have a
hammer
– sub-tasks might interact with each other
• if I am holding a hammer and nail, I can’t hold the boards
– so planning is a combinatorial problem
Intelligent Agents
• agents are: 1) autonomous, 2) situated in an
environment they can change, 3) goal-oriented
• agents focus on decision making
• incorporate sensing, reasoning, planning
– sense-decide-act loop
• rational agents try to maximize a utility
function (rewards, costs)
goals
perception
KB
initial state
action
goal state
agent
environment
• agents often interact in multi-agent systems
– collaborative
•
•
•
•
teamwork, task distribution
information sharing/integration
contract networks
voting
– competitive
•
•
•
•
•
agents will maximize self utility in open systems
negotiation
auctions, bidding
game theory
design mechanisms where there is incentive to
cooperate
Core Concepts in AI
• “Symbolic Systems Hypothesis”
– intelligence can be modeled as manipulating symbols
representing discrete concepts
• like Boolean variables for CupEmpty, Raining, LightsOn,
PowerLow, CheckmateInOneMove, PedestrianInPath...
• inference and decision-making comes from comparing
symbols and producing new symbols
– Herbert Simon, Allan Newell (CMU, 1970s)
• (A competing idea: Connectionism)
– neural networks
– maybe knowledge can’t be represented by discrete
concepts, but is derived from associations and their
strengths
– good model for perception, motor skills, and learning
Core Concepts in AI
• Search
– everything in AI boils down to discrete search
– search space: different possible actions
branch out from initial state
– finding a goal
• weak methods: depth-first search (DFS), breadthfirst search (BFS), constraint satisfaction (CSP)
• strong methods: use ‘heuristics’, A* search
S0
goal nodes
• Applications of Search
– game search (tic-tac-toe, chess)
– planning
– natural language parsing
– learning (search for logical rules that describe
all the positive examples and no negative
examples by adding/subtracting antecedents)
Core Concepts in AI
• Knowledge-representation
– attempt to capture expertise of human experts
– build knowledge-based systems, more powerful
than just algorithms and code
– “In the knowledge lies the power” (Ed
Feigenbaum, Turing Award: 1994 )
– first-order logic
• p vegetarian(p)↔(f eats(p,f)m meat(m)contains(f,m))
• x,y eat(joe,x)contains(x,y)fruit(y)vegetable(y)
•  vegetarian(joe)
– inference algorithms
• satisfiability, entailment, modus ponens, backwardchaining, unification, resolution
• Expert Systems
– Medical diagnosis: rules for linking symptoms
with diseases, from interviews with doctors
– Financial analysis: rules for evaluating credit
score, solvency of company, equity-to-debt
ratio, sales trends, barriers to entry
– Tutoring – rules for interpreting what a student
did wrong on a problem and why, taxonomy of
possible mis-conceptions
– Science – rules for interpreting chemical
structures from mass-spectrometry data, rules
for interpreting well logs and finding oil
• Major problem with many expert systems:
brittleness
• Major issue in AI today: Uncertainty
– using fuzzy logic for concepts like “good
management team”
– statistics: conditional probability that a patient
has meningitis given they have a stiff neck
– Markov Decision Problems: making decisions
based on probabilities and payoffs of possible
outcomes
Traditional Sub-areas within AI
• Natural language processing
– parsing sentences, representing meaning, metaphor,
answering questions, translation, dialog systems
• Vision
– cameraimagescorners/edges/surfaces
objectsstate description
– occlusion, shading, textures, face recognition,
stereo(3D), motion(video)
• Robotics – configuration/motion planning
• Machine Learning
– machines can adapt!
– learn from experience
– extract patterns from examples
– algorithms for: decision trees, neural networks,
linear classifiers...