From NARS to a Thinking Machine

1
From NARS to a Thinking Machine

• Pei Wang
• Temple University

2
Content

• NARS (Non-Axiomatic Reasoning System) a
  project aimed at building a general-purpose
  intelligent system, or a thinking machine
• The main ideas behind the project
• The development plan of the project
• The past, present, and future of the project

3
Observations

• Intelligence is a capability possessed by human
  beings, but not by animals and ordinary
  computers.
• The major difference not in what it can do,
  but in what it can learn to do.
• Key features
  adaptivity, generality, creativity, flexibility,
  but not absolute optimity.

4
Methodology

• Minimalism not to maximize the systems
  performance, but to minimize its theoretical
  assumptions and technical instruments, while
  still achieving desired performance.
• There are scientific and engineering reasons for
  following such an approach.
• Many such attempts have failed, but they might
  have followed wrong ideas.

5
Basic Principle

• Intelligence is the capability of a system to
  adapt to its environment and to work with
  insufficient knowledge and resources.
• The system should
• rely on constant processing capacity,
• work in real time,
• open to unexpected tasks,
• learn from experience.

6
Framework

• NARS is built within the framework of a reasoning
  system, with a language for knowledge
  representation, a semantics of the language, a
  set of inference rules, a memory architecture, a
  control mechanism.
• Advantages being domain-independent, combining
  the justifiability of individual steps and the
  flexibility of processes.

7
Categorical Language

• A typical sentence
• bird ? animal 1.0, 0.9
• Term bird and animal are names of concepts
• Inheritance relation ? special-general
• Truth value frequency, confidence

8
Experience-Grounded Semantics

• The truth value of a sentence is determined by
  available evidence in the experience
• f w/w, c w/(w1)
• Truth value uniformly represents randomness,
  fuzziness, and ignorance.
• The meaning of a term is defined by its
  experienced relations with other terms.

9
Basic Inference Rules
abduction
revision
10
Memory as a Belief Network
Cbird
11
Control Strategy

• In each step, a task is processed by interacting
  with a belief, according to certain rules.
• The task and belief are selected
  probabilistically, according to priority
  distributions among related tasks and beliefs.
• Factors influence the priority of an item
  quality of the item, usefulness of the item in
  history, and relevance of the item to the current
  context.

12
Compound Terms

• Compound terms sets, intersections, differences,
  products, and images.
• Variants of the inheritance relation similarity,
  instance, and property.
• New inference rules are added to carry out
  compound composition and decomposition.
• Related changes in memory and control.

13
Higher-Order Reasoning

• Two higher-order relations, implication and
  equivalence, are defined between statements.
• Compound statements negations, conjunctions, and
  disjunctions.
• The implication relation is used to carry out
  conditional and hypothetical inferences.
• Variable terms are used to carry out general and
  abstract inferences.

14
Procedural Reasoning

• Events as statements with temporal relations
  (sequential and parallel). Prediction and
  explanation as temporal inferences.
• Operations as statements with procedural
  interpretation. Skill learning and planning as
  procedural inferences.
• Goals as statements to be realized. Decision
  making as the making of new goals.

15
Development Progress

• DONE language definition
• semantics specification
• basic inference rules
• rules for compound terms
• rules for higher-order
  inference
• basic memory and control
• DOING rules for temporal/procedural inference
• refined memory and control

16
NARS Plus

• Optional extensions of NARS
• sensorimotor interface
• natural language interface
• education procedure
• socialization procedure
• special hardware
• evolution process

17
Conclusions

• An AI system should follow the same principles as
  the human mind, though it may have different
  internal structure, external behavior, practical
  ability, etc.
• To see intelligence as adaptation with
  insufficiency explains mental processes, guides
  system design, and distinguishes AI from other
  disciplines.

18
Information about NARS

• Website containing 30 publications and on-line
  demonstrations (a Java Applet and a Prolog
  program) of NARS (Version 4.2).
  (http//www.cogsci.indiana.edu/farg/peiwang/papers
  .html)
• Book Rigid Flexibility The Logic of
  Intelligence, Springer, ISBN 1402050445,
  Available September 15, 2006.
  (Wang-Contents-Preface.pdf)