Daily Note: Believable Agents

These notes are a summary of concepts presented in “Personality-rich believable agents that use language.”

A. Bryan Loyall and Joseph Bates. 1997. Personality-rich believable agents that use language. In Proceedings of the first international conference on Autonomous agents (AGENTS ’97). Association for Computing Machinery, New York, NY, USA, 106–113. https://doi.org/10.1145/267658.267681

  1. Core Characteristics
    • Believable agents
      • Autonomous agents with rich, specific personalities akin to characters in movies or animations
      • Integration of text generation with action, perception, inference, and emotion
      • Suspension of disbelief, creating a sense of reality for the viewer or user
  2. Communication and Expression
    • Use of language and action together for communication goals
    • Linguistic variation based on emotional state and personality
    • Real-time delivery of language with naturalistic elements like pauses and restarts
  3. Artistic Abstraction and Detail
    • Focus on essential traits for believability through abstraction
    • Importance of fine details (e.g., eye use, speech pauses, body awareness)
  4. Parallel and Incremental Behavior
    • Language generation occurs alongside other independent goals
    • Streams of control signals to multiple channels (e.g., eyes, body, voice)
    • Responsiveness to real-time events in the world
  5. Role of Emotion and Perception
    • Emotion linked to the success/failure of communication and other actions
    • Perception used to influence linguistic choices
  6. Behavior-Based Programming
    • Programs as collections of behaviors with goals, parameters, bodies, and preconditions
    • Hierarchical goal structures enabling parallel threads of action
    • Multiple top-level goals, reactivity, and backtracking mechanisms
  7. Components and Functions
    • Functions supporting emotion, social behavior, and memory inherit reactivity and parallelism
    • Hierarchical decomposition for structured concept generation
  8. Groups and Features Representation
    • Groups as roles with collections of features (name/value pairs)
    • Ease of creation and access to components
    • Central subgroup or feature (projector) guides group generation
  9. Processing and Communication
    • Combination rules for recent string generation
    • Communication between subgoals via parameter passing or dynamic variables
    • Dynamic scoping enables parallel behaviors to share information
  10. Inference and Context Sensitivity
    • Inference rules evaluated during goal pursuit
    • Sensor queries for real-time world interaction during active behaviors
    • Contextual annotations for adaptive behavior adjustments
  11. Real-Time Responsiveness
    • Flow rules ensure smooth information flow but must avoid excessive delays
    • Parallel behaviors react to real-time changes effectively