| Home | Table of Contents |
General Intelligence :
The planner searches in a state space in order to reach certain goals. State transitions are defined by actions. The triple of ( goals, states, actions ) defines a planning problem.
The goals of a problem can be explicitly passed to Working Memory. For example, a text query gets translated by the Natural Language Processor and passed to Working Memory.
The states of a problem may be directly observed from Working Memory, or alternatively, they can be obtained by asking "what-if" questions to the Inference Engine.
The actions of a problem consist of the preconditions and their effects. The planner may ask the Inference Engine "what can be done" in certain situations. Then it can ask "what if I perform this action" to obtain the effects of an action. Alternatively it can also perform an action to learn its effects reactively.
In the current "blackboard" architecture, the planner have direct access to memory systems (Semantic, Generic, and Episodic Memory). This enables the planner to do cognitive-level planning. For example "Try to recall the value of 13×13. If recall fails, use a calculator."
The ability of the planner to perform cognitive functions may be redundant to functions of the Inference Engine, but it is an advantage to have this redundancy because standard inference may be overriden when doing so is desirable.
The planner can learn to do things and store the acquired knowledge in Procedural Memory, which is private to the planner module.
The "chunking" learning method of Soar seems to belong to this category.
You may provide some instant feedback with this form:
| Home | Table of Contents |
Sep/2006 (C) GIRG [ Notice about intellectual property on this page ]