Kinds of AI agent structures

AI agents should have at least four out of five of the following characteristics:

• role
  • essence: specialization, style, tone
• memory
  • context: knowledge base provided to the agent
• reasoning
  • uses a model to decide an action
• actions
  • changes the environment
• learning
  • monitors feedback and adapts its functionality

Agents come in the following types:

• simple reflex
• model-based reflex
• goal-based
• utility-based
• learning

...but not all of the above are AI agents!

Simple reflex (non-AI) agents

• perceive the environment
  • take on a role to do something
  • uses a simple sensor that looks at least one variable
    • e.g. a thermostat
      • adjusts temperature by starting or stopping a heater
• a simple reflex agent however does not qualify as an AI agent
  • only has a role of performing some simple actions
  • no memory, reasoning nor learning
  • i.e. a simple reflex agent is a "simple" computer program (at best)

Model-based reflex agents

• perceive the environment and
  • take on a role to do something
  • know past actions with memory
    • e.g. a robot vacuum
      • builds a map of where it has cleaned
      • reasons where to clean next
      • takes actions to clean
• thus considered as primitive AI agents
  • they however function through fixed rules without learning
    • i.e. to focus on a part of the environment
    • e.g. a vacuum cleaner learning to clean dirtier areas more thoroughly

Goal-based agents

• not only perceive but adjust to a changing environment
  • re-evaluates the actions they need to take to achieve a goal
    • e.g. a navigational system
      • picks up new obstacles such as increased traffic volume and accidents
      • makes new routes based on new information
• thus, a higher form of AI agent than the model-based reflex agents

Utility-based agents

• instead of just reaching a goal
  • they try to do so with minimal friction
    • e.g. an investment agent
      • avoids volatile investments to achieve the safest return
• thus, a higher form than the goal-based agent

Learning agents

• have the ability to perform actions in unfamiliar environments
  • make sense of new data with existing data
    • e.g. a recommendation agent for movies / travel / etc.
      • tries to monitor an unfamiliar environment (us!)
      • learns to make recommendations based on learning via our "random" choices
• the highest form of AI agent (so far!)