Chapter 6: Building Control Algorithms for State Space Search

1
Chapter 6 Building Control Algorithms for State
Space Search

• Because we skipped ahead to Prolog, you should
  find everything in the first two sections of this
  chapter to be quite familiar
• Please review these sections on your own
• In essence, Prolog uses recursion and
  backtracking to control and implement a state
  space search through a space of logical
  inferences
• The other control and implementation approaches
  discussed in Chapter 6 are the production system
  and the blackboard architecture

2
Production Systems

• The production system is a computational model
  that has been used both to model human problem
  solving and to implement search algorithms
• Many expert systems are implemented as production
  systems
• Language translation, for both human and computer
  languages, can also be implemented using this
  approach
• A production system has three main components
• a set of production rules, each of which is a
  condition-action pair
• the condition tells when the rule is applicable
• the action tells what to do when the rule applies
• each production rule defines one problem solving
  step
• working memory
• this is a description of the state of the world
  (or at least, the current state of the problem
  being solved)

3
Parts of a Production System, continued

• The production rules interact with the working
  memory as follows
• When a condition of a rule matches the state of
  the world, then that rule can fire (ie., its
  action may be performed)
• When a rule fires, it can change the state of the
  world
• The third component of a production system is the
  recognize-act cycle
• Working memory, or the state of the world, is
  initialized to whatever is given by the problem
  description
• the state of the world is represented by a set of
  patterns
• All rules with conditions matching the patterns
  in working memory become a conflict set
• one rule is selected from the set, using some
  selection strategy, called conflict resolution
• the selected rule is fired
• The cycle repeats until nothing in working memory
  matches any rule conditions

4
Overview of a Production System
5
A Production System for Sorting abc Strings
6
Background of Production Systems

• Newell and Simon gave people problems to solve
  and then monitored the people as they solved the
  problems
• They recorded what the people said they were
  doing while they were solving problems, and also
  what they could see the people doing, such as
  moving their eyes from place to place
• These recordings were represented as a problem
  behavior graph
• A production system was then used to search this
  graph
• Production rules represented human problem
  solving skills (long term memory)
• The working memory was whatever the subject was
  paying attention to at a particular point in the
  problem solving process (short term memory)
• When this focus of attention matched the
  condition of a production rule, the persons
  focus of attention would change to the next thing
• Production systems, in general, may or may not
  embody such human problem solving strategies

7
The 8-Puzzle as a Production System
8
Search Space for 8-Puzzle with Depth Bound of 5
9
The Knights Tour Revisited
10
The Knights Tour as a Production System

• When we looked at the Knights Tour in Prolog, we
  wrote rules for each of the possible moves on the
  3 x 3 chessboard
• Ex move(1,8).
• This could be expressed as a production rule,
• condition the knight is on square 1
• action move the knight to square 8,
• or, using the ? notation for productions,
• knight on square 1 ? knight on square
  8
• To extend the approach to an 8 x 8 chessboard,
  you would not want to enumerate every possible
  move
• Rather, there are eight types of moves possible,
  leading to just eight production rules
• Ex condition current row lt 6 and current
  column lt 7
• action new row current row 2 and
• new
  column current column 1

11
Controlling Search in Production Systems

• To control search in production systems, you can
  choose data-driven or goal-driven search
  strategies, choose the structure of the rules,
  and choose the strategies for resolving conflicts
  (picking the rule to apply when more than one
  rules condition is satisfied)
• In the data-driven (forward chaining) approach
• Known facts are placed in working memory
• The recognize-act cycle finds the conditions of
  rules that match the fact patterns in working
  memory
• A rule with a matching condition is fired
• This creates new facts to add to working memory
• When a goal is added to working memory,
  computation stops
• When the rules are formulated as logical
  assertions, firing a rule corresponds to applying
  modus ponens to infer new facts

12
Example of Data-Driven Search

• In this example, the conflict resolution strategy
  is to choose the rule that has fired least
  recently

13
Goal-Driven Search

• To use a production system for goal-driven search
  (backward chaining)
• Start out with the goal you want to satisfy in
  working memory
• Instead of matching conditions and then firing
  actions, you match actions (which are
  conclusions) and choose a rule with a matching
  action from the conflict resolution set
• When you fire the selected rule, its conditions
  are added to working memory as new subgoals
• The cycle ends when you come to facts (which may
  be viewed as rules having actions but no
  conditions) justifying all of the subgoals in
  working memory
• These facts may be built into the program or may
  be asked of the user

14
Example of Goal-Driven Search (page 182)

• In this example, the start state is considered
  to be a known fact

15
Other Methods of Controlling Search in Production
Systems

• In predicate calculus, there is no implied order
  to truth
• In a production system, even though the
  conditions of several rules may all be true, the
  rules are fired in a particular order
• Ordering the rules affects the way in which
  search will proceed
• Ex In Prolog, the order in which rules is
  written is important
• Conflict resolution is a related issue
• You dont have to always fire the first true
  rule. Other strategies are
• Refraction This discourages cycles by preventing
  a rule from firing a second time unless the
  patterns in working memory that matched its
  conditions have changed
• Recency This keeps the search moving in one
  direction by choosing rules whose conditions are
  matched by the facts most recently added to
  working memory
• Specificity This prefers rules which match more
  specific conditions to those which match more
  generally

16
Advantages of Production Systems

• Production systems are used widely in AI
  applications because
• Knowledge and control are separated
• Problem solving knowledge is in production rules
• Control is in the recognize-act cycle
• They make it easy to implement state space search
• The state of working memory is a node in the
  graph
• The rules tell the allowable transitions between
  nodes
• Conflict resolution selects the branch of the
  graph to follow
• Production rules are modular (ie., very small
  independent modules)
• They can be added, deleted and modified
  independently of other rules (to some extent)
• Control is pattern directed
• What happens next is controlled by the state of
  the world

17
More Advantages

• The order in which rules fire may provide a
  solution path through a graph, and/or a trace of
  the way a person would solve a problem
• You can build a production system in almost any
  programming language, as long as you can match
  patterns
• Production systems may be used to model human
  cognition as well as to build practical systems
• Modeling human problem solving was Newell and
  Simons original intent for production systems

18
The Blackboard Architecture

• The blackboard architecture may be considered as
  an extension of the production system, although
  its possible to use this architecture with other
  reasoning mechanisms, as well
• The blackboard architecture is another approach
  to control in AI systems
• In this architecture, there are multiple problem
  solving modules, each of which is capable of
  solving some part of the problem
• The modules communicate with each other through a
  globally accessible structure called the
  blackboard
• The blackboard is (in a loose sense) a database
• Independent problem solving modules, called
  knowledge sources, regularly check the blackboard
  to see if there is data they can use
• If there is, they process the data, and post the
  results to the blackboard, providing additional
  data for any knowledge source to use

19
The HEARSAY Project

• The blackboard architecture was introduced by the
  HEARSAY project, a speech understanding program
  that interpreted and answered spoken questions
  about a library database of computer science
  articles
• Understanding speech is complex, and many factors
  enter into it
• Asynchronous, data-driven knowledge sources were
  built to analyze
• the waveform of the acoustic signal
• the phonemes of the acoustic signal
• the syllables these phonemes could produce
• the possible words these syllables could comprise
• the possible sequences these words could comprise
• the possible phrases these sequences could
  comprise
• Knowledge sources work on different parts of a
  problem and/or provide different possible
  interpretations for the same part of a problem
• This project led to the use of blackboards in
  many expert systems