Transcript Overview of Artificial Intelligence

Overview of
Artificial Intelligence
Thomas R. Ioerger
Associate Professor
Department of Computer Science
Texas A&M University
What is AI?
• Real applications, not science fiction
– Control systems, diagnosis systems, games,
interactive animations, combat simulations,
manufacturing scheduling, transportation logistics,
financial analysis, computer-aided tutoring, searchand-rescue robots
Different Perspectives
• Philosophical perspective
– What is the nature of “intelligence”? Can a
machine/program ever be truly “intelligent”?
– Strong AI hypothesis: Is acting intelligently sufficient?
– laws of thought; rational (ideal) decision-making
• Socrates is a man; men are mortal; therefore, Socrates is
mortal
• Psychological perspective
– What is the nature of “human intelligence”?
– Cognitive science – concept representations, internal
world model, information processing metaphor
– role of ST/LT memory? visualization? emotions?
analogy? creativity?
– build programs to simulate inference, learning...
• Mathematical perspective
–
–
–
–
–
Is “intelligence” a computable function?
input: world state, output: actions
Can intelligence be systematized? (Leibnitz)
just a matter of having enough rules?
higher-order logics for belief, self-reference
• Engineering (pragmatic) perspective
– AI helps build complex systems that solve difficult realworld problems
sense
– decision-making (agents)
– use knowledge-based systems
decide
act
to encode “expertise” (chess,
medicine, aircraft engines...)
weak methods:
Search
Planning
strong methods:
Inference
Search Algorithms
•
•
•
•
Define state representation
Define operators (fn: stateneighbor states)
Define goal (criteria)
Given initial state (S0), generate state space
S0
Many problems can be modeled as search
• tic-tac-toe
– states=boards, operator=moves
• symbolic integration
– states=equations, opers=algebraic manipulations
• class schedule
– states=partial schedule, opers=add/remove class
• rock band tour (traveling salesman problem)
– states=order of cities to visit, opers=swap order
• robot-motion planning
– states=robot configuration, opers=joint bending
1
Depth-first search
(DFS)
2
3
4
5
7
9
12
6
8
13
10
11
14
15
Notes:
recursive algorithms using stacks or queues
BFS often out-performs, due to memory limits for large spaces
choice depends on complexity analysis: consider exponential tree size O(bd)
1
Breadth-first search
(BFS)
2
5
6
7
8
17
18
19
20
14
15
16
3
9
4
10
11
12
13
Heuristics
• give guidance to search in terms of which nodes
look “closest to the goal”
– node evaluation function
– h(n)=w1*(piece_differential)+w2*(center_control)+
w3*(#pieces_can_be_taken)+w4*(#kings)
• greedy algorithms search these nodes first
• bias direction of search to explore “best” parts of
state space (most likely to contain goal)
• A* algorithm
– optimal (under certain conditions)
– finds shortest path to a goal
– insensitive to errors in heuristic function
Specialized Search Algorithms
• Game-playing
– two-player zero-sum games (alternate moves)
– minimax algorithm: form of “look-ahead” – If I make a
move, how will opponent likely respond? Which
move leads to highest assured payoff?
• Constraint-satisfaction problems (CSPs)
– state=partial variable assignment
– goal find assignment that satisfies constraints
– algorithms use back-tracking, constraint propagation,
and heuristics
– pre-process constraint-graph to make more efficient
– examples: map-coloring, propositional satisfiability,
server configuration
CSP algorithms
operate on the
constraint graph
•
Variables WA, NT, Q, NSW, V, SA, T
•
Domains Di = {red,green,blue}
•
Constraints: adjacent regions must have
different colors, e.g., WA ≠ NT
Planning
• How to transform world state to achieve goal?
• operators represent actions
– encode pre-conditions and effects in logic
pre-conds:
mixed(dry_ingr)&
mixed(wet_ingr)
Initial state:
in(kitchen)
have(eggs)
have(flour)
have(sugar)
have(pan)
~have(cake)
goto kitchen
mix dry
transfer
ingredients
ingredients
sautee
from bowl
to pan
bake at 350
buy
milk apply start
car
mix wet
frosting
ingredients
goto store
pre-conds:
x ingredient(x,cake)
dry(x)have(x)
effect:
mixed(dry_ingr)
Goal:
have(cake)
pre-cond: baked
another example to think about:
planning rescue mission at disaster site
Planning
• How to transform world state to achieve goal?
• operators represent actions
– encode pre-conditions and effects in logic
pre-conds:
mixed(dry_ingr)&
mixed(wet_ingr)
Initial state:
in(kitchen)
have(eggs)
have(flour)
have(sugar)
have(pan)
~have(cake)
goto kitchen
mix dry
transfer
ingredients
ingredients
sautee
from bowl
to pan
bake at 350
buy
milk apply start
car
mix wet
frosting
ingredients
goto store
pre-conds:
x ingredient(x,cake)
dry(x)have(x)
effect:
mixed(dry_ingr)
Goal:
have(cake)
pre-cond: baked
another example to think about:
planning rescue mission at disaster site
Planning Algorithms
• State-space search
– search for sequence of actions
– very inefficient
• Goal regression
have(cake) <= baked(cake)&have(frosting) <=...
– work backwards from goal
– identify actions relevant to goal; make sub-goals
• Partial-order planning
– treat plan as a graph among actions
– add links representing dependencies
• GraphPlan algorithm
– keep track of sets of achievable states; more efficient
• SatPlan algorithm
– model as a satisfiability problem
Knowledge-Based Methods
• need: representation for search heuristics and planning
operators
• need expertise to produce expert problem-solving behavior
• first-order logic – a formal language for representing
knowledge
• rules, constraints, facts, associations, strategies...
–
–
–
–
rain(today)wet(road)
feverinfection
in(class_C_air_space)reduce(air_speed,150kts)
can(take_opp_queen,X)&~losing_move(X)do(X)
• use knowledge base (KB) to infer what to do
– goals & initial_state & KB
do(action)
– need inference algorithms to derive what is entailed
• declarative vs. procedural programming
First-Order Logic
• lingua franca of AI
• syntax
– predicates (relations): author(Candide,Voltaire)
– connectives: & (and), v (or), ~ (not),  (implies)
– quantified variables: X person(X)Y mother(X,Y)
• Ontologies – systems of concepts for writing KBs
– categories of stuff (solids, fluids, living, mammals, food,
equipment...) and their properties
– places (in), part_of, measures (volume)
– domain-dependent: authorship, ambush, infection...
– time, action, processes (Situation Calculus, Event Logic)
– beliefs, commitments
• issues: granularity, consistency, expressiveness
D
Inference Algorithms
A&BD
• Natural deduction
B
– search for proof of query
A
BvC ~C
– use rules like modus ponens (from A and AB, get B)
• Backward-chaining
– start with goal, reduce to sub-goals
– complete only for definite-clause KBs (rules with
conjunctive antecedents)
• Resolution Theorem-proving
–
–
–
–
convert all rules to clauses (disjunctions)
{AvB,~BvC}AvC
keeping resolving clauses till produce empty clause
complete for all FOL KBs
Prolog and Expert Systems
• Automated deduction systems
• programming = writing rules
• make query, system responds with true/false
plus variable bindings
• inference algorithm based on backward-chaining
Prolog example
sibling(X,Y) :- parent(Z,X), parent(Z,Y).
grandfather(X,Y) :- father(X,Z),parent(Z,Y).
parent(X,Y) :- father(X,Y).
parent(X,Y) :- mother(X,Y).
mother(tracy, sally).
father(bill, sally).
father(bill, erica).
father(mike, bill).
?- sibling(sally,erica).
Yes
?- grandfather(sally,X).
grandfather(sally,mike)
• Unification Algorithm
– determine variable bindings to match antecedents of
rules with facts
– unif. algorithm traverses syntax tree of expressions
– P(X,f(Y),Y) matches P(a,f(b),b) if {X/a,Y/b}
–
–
P
P
X f Y
a f b
Y
b
also matches P(a,f(a),a)
does not match P(a,b,c), P(b,b,b)
• Managing Uncertainty in real expert systems
–
–
–
–
default/non-monotonic logics (assumptions)
certainty factors (degrees of beliefs)
probabilistic logics
Bayesian networks (causal influences)
• Complexity of inference?
– suitable for real-time applications?
Application of Data Structures and
Algorithms in AI
•
•
•
•
•
priority queues in search algorithms
recursion in search algorithms
shortest-path algorithm for planning/robotics
hash tables for indexing rules by predicate in KBS
dynamic programming to improve efficiency of
theorem-provers (caching intermediate inferences)
• graph algorithms for constraint-satisfaction
problems (arc-consistency)
• complexity analysis to select search algorithm
based on branching factor and depth of solution for
a given problem
Use of AI in Research
• intelligent agents for flight simulation
– collaboration with Dr. John Valasek (Aerospace Eng.)
– goal: on-board decision-making without ATC
– approach: use 1) multi-agent negotiation, 2)
reinforcement learning
• pattern recognition in protein crystallography
– collaboration with Dr. James Sacchettini (Biochem.)
– goal: automate determination of protein structures
from electron density maps
– approach: extract features representing local 3D
patterns of electron density and use to recognize
amino acids and build
– uses neural nets, and heuristics encoding knowledge
of typical protein conformations and contacts
• TAMU courses on AI
– CPSC 420/625 – Artificial Intelligence
– undergrad
• CPSC 452 – Robotics and Spatial Intelligence
• also related: CPSC 436 (HCI) and CPSC 470 (IR)
– graduate
•
•
•
•
•
•
•
•
•
•
•
•
CPSC 609 - AI Approaches to Software Engineering*
CPSC 631 – Agents/Programming Environments for AI
CPSC 632 - Expert Systems*
CPSC 633 - Machine Learning
CPSC 634 Intelligent User Interfaces
CPSC 636 - Neural Networks
CPSC 639 - Fuzzy Logic and Intelligent Systems
CPSC 643 Seminar in Intelligent Systems and Robotics
CPSC 644 - Cortical Networks
CPSC 666 – Statistical Pattern Recognition (not official yet)
Special Topics courses (CPSC 689)...
* = not actively taught
goals
perception
KB
initial state
action
goal state
agent
environment