Learning

One of the fundamental capabilities of intelligent agents is the ability to modify its behavior based on experience, to learn. It is generally accepted that any architecture which attempts to behave intelligently must have some learning component. However, the nature of learning supported by the architecture varies widely, from minimal support to a primary mechanism.

Learning and planning generally work together to a large extent, with planning generating opportunities for learning and learning helping to generate more efficient plans. An important condition for planning from experience is the ability to represent the world in symbolic form, in order to identify important features of the evironment and reason about them.

There are many different types of learning which have been observed in human behavior, and each architecture listed below attempts to emulate one or more of them. The following list provides a general description of the major identifiable learning types.

Explanation-Based Learning
EBL is a technique for capturing the results of processing performed by the agent into a compact representation for efficient retrieval the next time the agent encounters a similar problem. Most systems implement EBL by backtracing through problem-solving, summarizing the trace, and generalizing the result so that it is applicable to other problems.
Case-Based Learning
This technique maintains a list of cases in which a particular event occurred, and tries to detect commonalities in the cases which led to the event. The commonalities can now be considered preconditions for the event, and planning can take advantage of this knowledge when its goal is to achieve the event.
Learning by Abstraction
Abstraction learning attempts to express states and operators at a higher level to simplify processing.
Learning by Instruction
Tutorial learning involves taking instructions from another agent and learning from them in order to accomplish similar tasks alone.
Learning by Analogy
This technique transforms solutions to problems in one domain to solutions to problems in another domain by discovering analogous states and operators in the two domains.
Experimentation
When knowledge is incomplete, experimentation takes place to discover new knowledge about a particular domain
The following list of architectures include learning components.
Back to List of Other Capablities