University of Washington
Department of Computer Science and Engineering
CSE573-Artificial Intelligence I
Autumn 1997

Midterm Review Notes
November 5, 1997

===============================================================
Main topics we covered

1.  Basic "what is AI" discussion. 

What makes something intelligent?

    Structural versus observational criteria
          The Turing test

    An agent-based approach
          Emphasis on IO behavior
          Requiring a success measure
              Rationality versus limited rationality

    Main schools of "AI thought"  (mainly from readings)
       Psychology 
          Newell and Simon, means-end analysis
          Behaviorism and cognitive psychology
       Mathematics 
          Mathematical logic, complexity results, decision theory
       Linguistics
          Natural language processing and knowledge representation

    Historical concepts
       Microworlds and puzzle domains
       Weak methods and strong methods
       Expert systems versus common-sense reasoning


2.  Intelligent Agents 

    Definition of an agent as a mapping from percepts to action(s)
       Performance measure;  optimality versus satisficing
    Rationality and ideal rationality
    Reflex (reactive) versus deliberative agents


3.  Problem Solving 

    Two phases
       Problem formulation (goal then problem)
       Problem solution


    Dimensions for designing agents:

      Russell/Norvig dimensions
        Single state versus multiple state
        Contingency
        Interleaving
        Exploration

      Another way to look at it
        Agent's sensing and effecting capabilities
        Environmental characteristics
        Reward characteristics


    The general graph-search paradigm
        the non-deterministic version
        the priority queue version

    Uninformed methods
        breadth first, uniform cost
        depth first, 
        depth-limited and iterative deepening
        regression and bidirectional search

    Informed methods
        definition of "heuristic search"
        the usual components of an evaluation function
            (state so far, estimated state to go)
        greedy search
        A* search, admissible heuristics, and graceful degradation
        IDA*

    Iterative improvement
        hill-climbing, gradient descent, simulated annealing

5.  Logical Reasoning  (Chapters 6, 7, 9)

    Problem-solving agents and the general KR paradigm

    Components of a logic:  syntax, inference, semantics

    Common logics:  
        propositional
        first-order
           with equality, sets, numbers

    Semantics:
        interpretations, models

    Properties:
       validity, unsatisfiability
       tautology, inconsistency
       soundness, completeness
        

    (SKIP 7.2 to the end of the chapter)
    (SKIP Chapter 8)

    Automated inference in first-order logic
      inference as a search task
         forward and backward chaining

    Resolution
      proof by refutation
      
      procedure
         convert database to clause form
            DB is a conjunction of disjuncts
            no implication
            no existential quantifiers (Skolem terms)
         negate goal, convert to clause form, and 
           add to database
         try to produce a derivation of the empty clause

      completeness of resolution

    Horn clauses and Prolog
      what is a Horn clause?
      what is the sacrifice in expressive power?
      what is the algorithmic gain?