Design of Computer Problem Solvers

1
Design of Computer Problem Solvers

• Introduction

2
What happens today?

• Syllabus
• Course mechanics and administriva
• Overview of course
• How do AI programs differ from conventional
  programs?
• What types of reasoners will we build?
• Homework 0
• Preview of Homework 1

3
What this course is about

• Building systems that reason
• By solving problems, both simple and complex
• By storing their dependencies and constraints
• By explaining their results
• How to build well-engineered, efficient AI
  systems
• Start with simple reasoner, and grow with task
  complexity
• Use industrial-strength AI programming

4
How are AI systems different from conventional
programs?

• All programs must maximize these constraints
• Efficiency How fast and small is the code?
• Coherence How well-tuned to its particular
  domain?
• Flexibility How adaptable to new domains or
  unusual circumstances?
• Additivity How easily can new info be added?
• Explicitness How clearly is knowledge
  represented?
• Which is more important? Depends.

5
How are AI systems different from conventional
programs?

• All programs must maximize these constraints
• Efficiency
• Coherence
• Flexibility
• Additivity
• Explicitness

Conventional programs
AI programs
6
Conventional programming approach
Large Hunk of Code
Resulting computer program
Knowledge of the domain Programming Knowledge
7
Example Traffic control programs

• Very efficient
• Handles routine situations well
• Doesnt work well in unusual circumstances
• Hard to extend or adapt to new circumstances
• Hard to figure out why it does everything it does
• Engine tester.

8
The essence of AI programming
Programming Knowledge
Reasoning Engine
Knowledge of the domain
Computer Program
Representation of Domain Knowledge
9
Efficiency Coherence Flexibility Additivity Expli
citness
ConventionalTradeoffs

• Efficiency Coherence improve together
• Store only the minimum necessary to get job done
  (reduces explicitness)
• Figure out optimal order of actions in advance
  (reduces flexibility)
• Both of these reduce additivity because knowledge
  resides only in the programmers head.

10
Efficiency Coherence Flexibility Additivity Expli
citness
AI Tradeoffs

• Flexibility, Additivity, and Explicitness improve
  together
• Explicit representation of knowledge allows
  program to figure out what to do (but with
  reduced coherence and efficiency).
• Program can be extended by adding new knowledge
  rather than mind surgery (but with reduced
  coherence and efficiency).

11
How you resolve these tradeoffsdepends on the
task

• How often will the program change?
• How fast is the processor? What are the
  real-time constraints?
• How cheap/available are programmers or
  content experts?

12
Reasoners of increasing level of sophistication

• Classical Problem Solvers, cfr. AIMA
• Pattern-directed inference system
• Truth maintenance systems

13
Classical problem solving

• In classical problem-solving
• Problems are defined as a search space of
  potential solutions
• Domain-independent search, e.g. BFS, DFS, , is
  used to find a problems solution
• Examples we explored
• Path planning
• Map coloring

14
Pattern-directed Rule Systems

• Programs are organized using a system of rules
• Rules contain patterns which represent
  antecedents, and allow the inference of
  consequences
• Examples we will explore
• Logical theorem prover (Natural Deduction, cfr.
  Grondslagen I)

15
Truth Maintenance Systems

• Adds a dependency network to pattern-directed
  rule system
• Allows reuse of dependencies, greatly increasing
  the efficiency of reasoning
• Types of TMSs
• Justification-based,
• Logic-based, and
• Assumption-based

16
Justification-based TMS

• Basic type of TMS
• Records a network of dependencies between Horn
  clauses
• Will be used to illustrate dependency-directed
  search
• Examples we will explore
• Symbolic integration (JSAINT)

17
Logic-based TMS

• Operates over logical values
• Utilizes Boolean constraint propagation (BCP) for
  quick inferencing
• Detects and handles contradictions
• More efficient for dependency-directed search
• Examples we will explore
• Cryptarithmetic puzzle solver

18
Assumption-based TMS

• Flexible handles multiple sets of assumptions
  (including possible inconsistent assumptions)
• Good for reasoning in multiple contexts at once
• Useful for diagnosis and planning
• Examples we will explore
• Blocks World

19
Prerequisites

• What youll need to know about
• Lisp, including the use of structs or classes
• How to write and use pattern matchers and
  unifiers, cfr. Abelson and Sussman The Structure
  and Interpretation of Computer Programs
• Some experience with knowledge representation

20
Good auxiliary texts for this course

• Steele, Guy. Common Lisp The Language. 2nd Ed.
• Norvig, Peter. Paradigms of Artificial
  Intelligence Programming Case Studies in Common
  Lisp.