Soar - Learning

Description

• The Soar architecture directly supports learning using its chunking mechanism.
• Learning does not occur simply for the sake of learning but rather it occurs as a direct result of problem solving. That is, learning occurs reflexively when impasses have been resolved.
• Thus learning serves to increase task knowledge by grouping individual steps of the task plan into larger steps called "chunks". This increased knowledge will decrease problem space search and thus increase the efficiency of performing the task.
• Chunking is intended to serve as a universal learning mechanism that can emulate other learning mechanisms by the addition of knowledge. Possible learning mechanisms include:
  • Learning of procedural knowledge may arise from the chunking of subgoals that involve problem space creation and selection.
  • Learning of episodic knowledge may arise from the chunking of subgoals that involve state creation and selection.
  • Learning of declararative knowledge may arise from the chunking of subgoals that involve operator creation, slection and execution.