CS G120 Artificial Intelligence

1
CS G120 Artificial Intelligence

• Prof. C. Hafner
• Class Notes Jan 8, 2009

2
Goals of artificial intelligence (AI) field

• Artificial systems with humanlike ability to
  understand and reason (cf. cognitive science)
• Solve problems that are too large to find the
  best answer algorithmically, so require
  incomplete methods
• Solve problems that are not well-understood

How do we judge whether we have
succeeded? getting the right answer ? the
Turing Test ? (and modified versions) usefulness
of the resulting techniques? we know it when we
see it?
3
Artificial systems with humanlike ability to
understand and reason

• Main techniques formal logic, knowledge
  representation, automated deduction (employs
  search algorithms)
• Uses problem-solving, natural language
  understanding, planning, intelligent HCI

4
Artificial systems with humanlike ability to
understand and reason (cont.)

• Main techniques evidential logics (probability,
  fuzzy logic, . . .), Bayesian inference nets,
  Markov models
• Uses Problem solving under uncertainty, decision
  support systems (expert systems)

5
Solve problems that are too large to find the
best answer algorithmically, so require
incomplete methods

• Main techniques heuristic search
  dependency-directed backtracking
• Uses production scheduling and other constraint
  satisfaction problems game-playing
• Uses a component in large-scale reasoning and
  planning systems

6
Solve problems that are not well-understood

• Main techniques weighted rule-based sytems
  bayesian inference nets statistical induction
  and machine learning in general
• Uses Finance, search engines, computational
  science (discovery), data mining
• Uses computer vision systems that see and
  recognize objects

7
Go over syllabus
8
Framework for Intelligent Agent Design

• What can the agent do? (primitive actions)
• What can the agent know?
• Input to the program
• Symbolic input from keyboard, files, networks
• Sensing of the physical world (most often
  simulated)
• Memory of its history
• World knowledge

9
Types of Agents

• Reflex agent
• no state or memory
• Reacts to current input according to its program
  (condition ? action rules)
• Knowledgeable agent
• Reacts to current input by understanding its
  meaning relative to a knowledge base
• Reacts according to rules, but the condition may
  be complex and require reasoning to verify

10
Types of Agents (cont.)

• Goal-based agent
• Condition ? action rules replaced by a
  goal-seeking framework
• One or more goals
• A set of operators that specify, for each
  action
• Conditions that need to be true before the action
  can be performed
• Effects of the action

11
General Problem Solver Example of an early
goal-based system (Herbert Simon won Nobel prize
in Economics)
Employed means-ends analysis (a pre-cursor of
backward chaining now used in many systems
Goal Transform situation A to situation B
suceed
Match A to B to find difference D
A
Subgoal Reduce D
Transform A into B
Success
none
fail
fail
Fail
Fail
Success
12
Means-ends analysis (cont)
Goal Reduce difference D between situations A
and B
A
Search for operator Q relevant to reducing D
Subgoal Apply Q to A producing A
Success
fail
fail
none
Fail
Fail
Goal Apply operator Q to A
A
A
Match A to the conditions of Q, finding
difference F
Apply Q to A
Subgoal Reduce F
Success
fail
fail
none
Fail
Fail
13
Agents and environments

• The agent function maps from percept histories to
  actions
• f P ? A
• The agent program runs on the physical
  architecture to produce f
• agent architecture program

14
Vacuum-cleaner world

• Percepts location and contents, e.g., A,Dirty
• Actions Left, Right, Suck, NoOp

15
Reflex agent

• Behavior depends only on current state (no
  history)
• Approach 1 Behavior expressed as a program
• Approach 2 Behavior expressed as a set of
  production rules (called table-driven by RN)
• Condition ? Action
• Condition ? Action
• . . .
• Condition called left-hand-side (LHS)
• Action called right-hand-side (RHS)

16
Simple Reflex Agent Vacuum Agent
Approach 1 (hard-coded behavior) if status
Dirty then return Suck else if location A then
return Right else if location B then return
Left
Called reflex because there is no world model
17
Simple reflex agents
18
Simple Reflex Agent Vacuum Agent
Approach 2 (rule-based) state ? perceive()
rule ? rule-match(state, rule
base) action ? RHS(rule) return action
Called reflex because there is no world model
19
Production rule system

• Drawbacks
• Huge RULE BASE (time consuming to build by hand)
• Inflexible (no adaptation or learning)
• What if more than one condition is satisfied?

20
Knowledgeable agents
21
Knowledge bases

• Knowledge base set of sentences in a formal
  language
  
• Declarative approach to building an agent (or
  other system)
• Tell it what it needs to know
• Then it can Ask itself what to do - answers
  should follow from the KB
  
• Agents can be viewed at the knowledge level
• i.e., what they know, regardless of how
  implemented
  
• Or at the implementation level
• i.e., data structures in KB and algorithms that
  manipulate them

22

• Q What formal language(s) can we use to
  represent
• Current facts about the state of the world
• General facts about how the world behaves
• General facts about the effects of actions
  thatthe agent can perform
• Condition ? action rules that specify how
  theagent will behave
• A Formal logic
• Syntax and semantics well understood
• Computational tractability known for
  importantsubsets (Horn clause logic)