Artificial Intelligence

1
Artificial Intelligence

• Artificial Intelligence (AI) is the name given to
  encoding intelligent or humanistic behaviors in
  computer software.
• Problem Nobody has created a widely accepted
  definition of intelligence.
• At one time was considered a uniquely human
  quality.
• Now generally accepted to be an animal quality.
• Has been linked to tool use, tool creation,
  learning, adaptation to novel situations,
  capacity for abstraction.
• Problem Nobody has created a widely accepted
  definition of artificial intelligence.
• Cognitive models attempt to recreate the actual
  processes of the human brain.
• Behavioral models attempt to produce behavior
  that is reasonable for a situation regardless of
  how the behavior was produced.
• Tend to focus on reasoning, behavior, learning,
  adaptation.

2
Artificial Intelligence Challenges

• Format of Knowledge the data structures we have
  discussed so far capture data values, but not
  data meaning.
• Graphs, trees, lists.
• Size of Knowledge How do you store it all?
  Once stored how do you access only the pertinent
  items and skip over irrelevant items.
• Humans are good at this, though we dont know
  why.
• Relationships between Pieces of Knowledge This
  is worse than the size of knowledge.
• Given n items and m types of relationships, there
  are m(n2) possible relationships.
• Is it better to explicitly represent
  relationships or derive them in real time as we
  need them?

3
Artificial Intelligence Challenges

• Ambiguity Knowledge ultimately represents
  natural phenomena that are inherently ambiguous.
  How do we resolve this?
• Acquiring Knowledge How does one combine new
  and old information?
• Relationship to old knowledge.
• Negative learning can we detect false
  information or contradictions?
• Can we quantify the reliability of the knowledge?
  Truth nets attempt to do this.
• Deriving Knowledge, Abstracting Knowledge Given
  a set of information, can I derive new
  information? Reasoning systems and proof systems
  attempt to do this. Can I group similar
  knowledge items into a more general single item?

4
Artificial Intelligence Challenges

• Adaptation How can I use what I know in new
  situations? What constitutes a new situation?
• Sensing Sensing is the ability to take in
  information from the world around you. Virtually
  all computer systems Sense 1s and 0s through
  keyboard, mouse, and serial port.
• Perception Perception is related to sensing, in
  that the meaning of the thing sensed is
  discovered. Auto example.
• Emotional Intelligence
• I think therefore I am. Renee Descartes, about
  1640.
• Descartes Error is a book by Antonio R Damasio,
  1995, in which he proposes that traditional
  rational thought without emotional content fails
  to create intelligent behavior.
• Social Knowledge, Ethics How do I behave with
  my teammates, strangers, friend, foe? What are
  my responsibilities towards others as well as
  myself?

5
Proposed AI Systems

• Rule Based Behavior designed behavior
  specifying sets of conditions and responses.
• Finite-State Machines Graphical representations
  of the state of systems, with sensory inputs
  leading to transitions from state to state.
• Scripts attempts to make behavior production
  tractable by anticipating behaviors that follow
  certain sequences. The Restaraunt Script is a
  typical example we expect roughly the same
  behaviors (be greeted, be seated, order drinks,
  get drinks, ) no matter what restaurant we are
  in.
• Case-based and Context-Based Reasoning attempt
  to reduce search space of possible behaviors by
  only considering those associated with certain
  situations or contexts.

6
Proposed AI Systems

• Cognitive Models Attempts to model cognitive
  processes.
• Cognitive Processes attempt to match human
  thinking by reproducing human thought processes.
• Neural Nets attempt to match human thinking by
  reproducing brain synapse structures.

7
Proposed AI Systems

• Emergent Behavior Overall behavior resulting
  from the interaction of smaller rule sets or
  individual agents. Overall behavior is not
  designed but desired.
• Genetic Algorithms represents behavioral rules
  as long strings, termed genomes. Behavior is
  evolved as various genomes are tried and
  evaluated. Higher rated genomes are allowed to
  survive and reproduce with other high ranking
  genomes.
• Ant Logic Named after the behavior of ant
  colonies, where individuals have very simple rule
  sets, but complex group behavior emerges through
  interactions.
• Synthetic Social Structures Models more complex
  animal social behaviors, such as those found in
  herds and packs. Allows efficient interaction
  without much communication.

8
(No Transcript)
9
Natural Language Vocal Interaction Between Live
and Synthetic Agents

• Keith Garfield and Donald A. Washburn
• Institute for Simulation Training
• University of Central Florida
• 407-882-1342, 407-882-1433
• kgarfiel_at_ist.ucf.edu, dwashbur_at_ist.ucf.edu

10
Agenda

• Problem Description and Motivation
• Issues associated with automated voice processing
• Natural Language Voice Interface (NLVI) Overview
• Summary

11
Problem Description and Motivation

• Overall Goal Produce a technology allowing
  natural language vocal interactions between live
  and virtual entities.
• Overall Motivation Enable
• Reduced staffing requirements appropriately.
• Virtual team members.
• Virtual trainers/coachers/advisors.

12
3 Phases of Automated Speech Processing
Speech to Text
Natural Language Processing
MIC
Text to Speech
SPKR
13
Speech-To-Text (STT) Processing

• Purpose Convert the spoken word to text.
• Techniques
• Match signal (digitized) to dictionary of sounds
  and words.
• Improve accuracy via syntactic analysis (not
  semantic).
• Improve accuracy by tracking history of the
  speaker.
• Challenges
• Differences in speech between persons/genders
• Differences in pronunciation given by the same
  person over time and in different situations.

14
Speech-To-Text (STT) Processing (cont)

• Notes
• Quite a bit of research currrently in this area
• Bell Labs
• Carnegie Mellon University (SPHINX)
• Commercial Products available and used
  successfully.
• VivaVoice
• DragonNaturallySpeaking
• MSSpeech API
• Speaker-dependent systems achieve about 95
  accuracy.
• Speaker-independent systems may have poor
  accuracy, rely on limited vocabularies.
• Future systems will probably be multi-modal
  (voice and gesture, voice and touch-screen)

15
Text-To-Speech (TTS) Processing

• Purpose Converting text to spoken word.
• Techniques
• Match text to phonetic dictionary of
  sounds/words.
• Incorporate emotional content by intonations as
  suggested by punctuation and context.
• Direct changes in pitch, volume, and speed to be
  imbedded explicitly in the text using special
  symbols (XML).
• Challenges
• Lack of models relating intonations to emotion or
  intent.
• Technically difficult to reproduce natural
  sounds.
• Notes
• Commercial product available not heavily
  researched by IST.

16
Natural Language Processing (NLP)

• Purpose (1) Extract meaning from text.
• Purpose (2) Compose text conveying a specific
  meaning.
• Techniques
• Parse sentences, often using Finite State Machine
  models ot tree-like data structures.
• Store meaning in a knowledge representation
  database, often rule-based or realtional
  database.
• Produce sentences using parse trees and
  semi-random word selection to compose sentences.

17
Natural Language Processing Challenges

• Parsing Sentences Syntax and semantics are
  intertwined in natural languages.
• Storing Meaning Knowledge representation is
  still a difficult problem. Number of rules and
  relationships required to cover non-trivial
  domains is large.
• Extracting Meaning from text
• Word meanings change when context changes.
• Idioms, metaphors, and similes provide
  challenges.
• Emotional content colors meaning (e.g. sarcasm or
  humor)

18
Natural Language Ambiguity

• Lexical Ambiguity - one word, many meanings
• Stay away from the bank.
• Structural Ambiguity - one sentence, differents
  grammatical structure.
• He saw that gasoline can explode.
• (Source Winograd, "Computer Software for Working
  with Language)

19
Natural Language Ambiguity

• Deep Structural Ambiguity - Same sentence, same
  grammatical structure, different meaning.
• The chickens are ready to eat.
• Semantic Ambiguity - Same phrase can have two
  meanings.
• David wants to marry a Norwegian.
• Pragmatic Ambiguity - confusin use of pronouns.
• She dropped a plate on the table and broke it.
• (Source Winograd, "Computer Software for Working
  with Language)

20
NLVI Overview Targets Military Trainers

• Immersive Command
• Environments Station

21
NLVI History

• Voice Federate (VF) project is a predecessor of
  NLVI.
• VF demonstrated the feasibility of allowing vocal
  control over synthetic entities in an immersive
  simulation.
• VF used SAICs Dismounted Infantry Semi-Automated
  Forces (DISAF) as a CGF platform.
• VF allowed basic command and control over CGF
  entities
• Allowable commands were a subset of existing
  DISAF unit and individual behaviors. No new
  behaviors created.
• Scripted synthetic speech was generated to
  confirm receipt of orders, give notice of task
  completions, and provide spot reports when enemy
  forces sighted.

22
Summary

• Automated voice processing can be broken down
  into Speech-To-Text, Natural Language Processing,
  and Text-to-Speech phases.
• The Natural Language Vocal Interface (NLVI) will
  allow limited natural language conversations
  between live and synthetic participants in
  simulations.
• NLVI makes use of the limited knowledge domain
  and formalized military speech to aid in voice
  processing.
• NLVI allows for queries to resolve ambiguities in
  the natural language processing.

23
NLVI Systems
24
WDB Components
Engine Receives information - Generates responses
- Updates STS/LTL/CBQ
Long Term Status (LTS) Echolon Structures Mission
Goals, Intel Terrain/Nav Info
Short Term Status (STS) Locations Current
task Visibility/Threats
Conditional Behavior Queue (CBQ) Misnomer A set
of possible behaviors From which to
choose ltpredicategtltbehaviorgt Ambiguous behaviors
generate info requests