Deliberative vs Reflexive Learning
Deliberative learning is when an architecture decides if it is
worth its while to learn something. The decision may manifest itself as
several smaller issues:
- When does learning take place?
- What is learned? What is ignored?
- Are learned rules kept forever, or may they be forgotten?
- How is learning done?
Some of these decisions are made when the new information first arrives, but
some made be made later. In particular,
Prodigy keeps running statistics on how well it is served by the control
rules that it learns. Balancing the decrease in problem space search cost
gained by additional knowledge against the increase in knowledge search cost
is known as the utility problem.
The advantages of deliberative learning are:
- memory savings for the most effective rules,
- time savings for matching rules
- the system designer has to do less work to make sure the agent learns only
what it is supposed to
Reflexive learning is when an architecture automatically keeps
all rules it creates. No statistics on rule efficiency are kept. The
advantages of reflexive learning are:
- simplier learning mechanisms,
- time and memory savings by not keeping statistics on, and grading, rules
Press this line for general discussion on learning.
Examples of deliberative learning architectures are:
Examples of reflexive learning architectures are:
Examples of "hybrid" learning architectures are:
Examples of architectures that make no commitment are:
Other Properties.
Back to the Title Page.