Communication Model 1 -

Notes for 19^th September 2001:

by Rajat Sharma

M – Memory

R – Reasoning

These types of Agents perceive and act. No Natural Language is needed. There is no verbal communication, only actions like pointing etc.

In these Agents, results of reasoning are encoded and sent. The receiver decodes the language.

These Agents use telepathy. This can be implemented using Shared Memory.

No explicit Encoding/Decoding is required. The agents are structured such that the results of reasoning is directly accesible

Blackboard Architecture -

As soon as one agent writes something to the shared memory, other agents can access it.

Ambiguity –

A word can mean different things to different agents.

For instance, "I saw the man on the hill with the telescope".

This sentence can have 3 meanings. This is an example of Structural Ambiguity.

There can also be Word-Sense Ambiguity, that is, different meanings of a word. Eg: coke.

There’s very little ambiguity in telepathy, since you know what the other person is thinking.

There is a lot of ambiguity in Communication Model 2, because they can communicate abstract concepts. Though there is lesser ambiguity in Communication Model 1 there is limited expression.

What is the internal representation?

Internal Representation refers to internal encoding that is available for manipulation.

Is the internal representation verbal or non-verbal?

We generally don’t think in English (or any other Natural Language)…

One way of thinking is that all languages arise from a Universal Grammar L_o (Chomsky).

_{L_ohas some rules
All other languages implement these rules in slightly different ways.

There is no internal representation in reactive control. Therefore we can use only CM1. (Communication Model 1)

For telepathy to work we need agents to have the same internal representation. We assume –

Universal Language for internal representation
Common Language for internal representation

CM1 –

good for certain tasks
not very complicated, less ambiguous

CM2 –

can have different internal representation
need only a common language (or encoding).

CM3 –

requires contract between agents as to what all is accessible to others

Theory of Natural Language Processing can't be identical to theory of Programming Languages. Lexical Analysis and Syntactic Analysis go together.

Eg: Syntactic Analysis -
S -> NP VP
NP -> noun | pronoun
VP -> verb [determinant] NP

Lexical Analysis -
He . . . . . . . . . pronoun
drank . . . . . . verb
whisky . . . . . noun

Lexical Analysis (LA) is hard because words can fall into multiple categories.

Syntactic Analysis (SA) is hard because the number of rules can be large, and rules may have choices.
Eg: VP -> verb [det] NP | verb [det] PP NP

LA becomes more complicated when number of mappings increases and choices increase.

SA becomes more complicated when number of rules increases and choices increase.

VISION –

In vision we have a labeling problem.
In NLP (Natural Language Processing), labeling problem is solved using rules (ie. Constraints).

Interpreting an image -

Labeling is needed to identify surfaces. For this we need to characterize edges, that is, identify edges that connect two surfaces.

I drank the vodka.
P V det N -> labels satisfying constraints.

The Constraint Satisfaction Problem –

Given: Multiple Constraints
Find: A set of labels that satisfy all the constraints.

There is no first stage of language because society has already made a way of distinguishing words. But it is not so on Vision, because there is no demarcation of edges.

It's hard to identify which arrow formed what part of the image.

In reading also, when there is a gap (space) we identify a word.}