Knowledge Structures

Lecture notes for CS8803B Artificial Intelligence, Fall 2002, 9/30/2002

Logic:

man(marcus)

Production Systems:

if “tawny color” and “dark spots” then “giraffe”

Semantic networks

Agents perceive actions (input) and produce output based on their memory.

Logic agent

Marcus à memory organized in logic statements à outputs man(marcus)

Information output increases with each type

(input) man(marcus)

Production System

System based on rules yields more information than logic agent could about

marcus

Semantic Networks

Yields more information than logic agent or production system about marcus

All of these types of agents share the fact that memory yields information based on queries. The advantage to structuring knowledge is that it is possible to yield more inferences than with unstructured knowledge. Are there other ways to structure knowledge, in order to elicit even more inferences?

Russell/Norvig approach to agents- agent types

Memory plays a bigger role with each agent

Reflexive

State based

Goal based

Utility based

With each successive type, memory plays a bigger role, and we can answer many more questions when structure is applied to memory.

Goal: develop other knowledge structures that elicit information with little input

Frames (objects)

Example: Develop a software agent that reads stories from the CNN website, and can answer questions about those stories.

CNN: “The president of Lower Slabovonia said that an earthquake occurred along the Sadiz fault line today, causing massive damage and numerous casualties.”

A human reader understands what occurred, where, when, and how serious the effects were from the event. If the agent has general knowledge about earthquake events in general, then it can first recognize the event from the story, and make expectations about more information on the event as it becomes known in the story. A frame structure helps this process.

Frame : a structure in which a large number of slots are combined under one name. Analogous to software engineering classes and objects

Slots	Fillers
Time	<unspecified>
Day	Today
Place	Lower Slabovonia
Fault	Sadiz
Damage	Massive
Casualties	Numerous

	Slots only apply to the earthquake event frame

…so now when the story is posted at CNN, the agent can pop up a new frame for an earthquake, and generate expectations about the facts that should appear in the rest of the story. These expectations are based on the slots that are enumerated in the earthquake frame. They are filled in as the story progresses.

Agent diagram

Observations

• There are many events in the real world (not just earthquakes), so is it necessary to generate a new frame for each possible event?

• How can we access the correct frame, given the large number of frame that must exist to describe the real world?

• Many events have slots in common, is it possible to structure frames to take advantage of this? (example- events all have a place and time, common attributes)

We can organize frames in a frame hierarchy (these ideas parallel object oriented concepts developed in software engineering contemporary to AI frames)

Events example diagram (ako in diagram = a kind of)

Observations

• Stereo typicality frames representing real world objects should contain slots that describe the stereotypical attributes of the of the objects (ie. a car frame should not describe its buoyancy)

• Inheritance – a lower level frame should have access to the higher level frame slots, provided that the lower level frame does not override them

Example: in the CNN story the president said that the earthquake originated along the Sadiz Fault Line. (the agent can now update the fault slot for the earthquake frame.)

Now we want to query the agent: How much damage in US$ is estimated to have occurred? The agent does not know about the damage statistics, because they have not been provided in the story.

Russel/Norvig: agent assumption that all information is available to the agent in the world.

Counter position: The world doesn’t always have all the information about everything that we want to know about, but the agent should be able to give a reasonable answer. By using the frame system, this is possible.

• Default Values can be supplied to slots so that inferences can still be made even though the true facts are not (yet) known.

Characterization of Concepts

What is a concept of a bird?

Representations:

If there is a bird that does not fly (i.e. Ostritch) then it would not pass this test

Logic:

Rules:

if “small” and “can fly” and “has feathers” and “has wings” then “bird”

Frames: Frame Hierarchy for birds

Observations on frame hierarchy

• Frames are not definitions, but rather a characterization of the attributes of a stereotypical object (bird)
• Does not have to be logically consistent
• Allows for easier methods to deal with exceptions to rules (Ostrich)

How are frames and semantic nets related?

Example: Atlanta Braves manager Bobby Cox likes pitcher Tom Glavine

Semantic Network Diagram

Frames Diagram

So frames are equivalent to semantic nets in terms of the information captured, and frames are also equivalent to prepositional logic in information captured.

Propositional Logic Representation

Each proposition is atomic, with no relation between propositions

Is-a(manager, (Bobby Cox))

Is-a(pitcher, (Tom Glavine))

Likes((Bobby Cox),(Tom Glavine))

The difference between the representations is the ways to search each of these representations for information. The goal of knowledge representation is to be able to do things with the knowledge once it is stored. Frames enable inferences and easier access. The knowledge contained in each representation is the same, but the methods to use that information vary.