%----------------------------------------------------------------------------

INTELLIGENT AGENTS
(largely from Russel and Norvig)

%----------------------------------------------------------------------------

An agent is capable of interaction with its environment, of perceiving
changes in it, and of effecting changes by acting on it.

The actions of an rational agent result in the achievement of its
goals --- this is called principle of rationality (according to
Newell).

A key characteristic of intelligent agents in autonomoy. 

%----------------------------------------------------------------------------

Agents have architectures. 

An architecture of an agent consists of a set of components and a set
of connections among the components. 

%----------------------------------------------------------------------------

Agents may or may not be physically embodied.

Biological animals (including humans) are examples of physically
embodied intelligent agents. Robots too are examples of the same.

Software agents (and, these days, network or internet agents) are
examples of agents that are not physically embodied. These are sometimes
called softbots (or knowbots) after robots.

%----------------------------------------------------------------------------

Agent Architecture 1: Reflexive Agent

Reflexive agents directly map percepts into actions.  These agents are
also called situated action agents or reactive control agents.


			Agent 				Environment


			Percepts      <--- Sensors      <----
		        |
			|
			|
		        V
		       State of 
		       the world
                        |
                        |
 condition-             |
 action rules           |
                        |
                        V
                       Actions		---> Effectors   --->


The condition-action rules are also called situation-action rules,
or production rules or, simply, productions 

Assumption: The world is perceptually accessible.

Advantages: simplicity of design; efficiency of processing 

Disadvantages: behavior can be very complex (if the world is complex),
the agent can fail (if the world is not perceptually accissble)

%----------------------------------------------------------------------------


Agent Architecture 2: State-Based Agent

State-based agents have internal states that store information 
about the history of the world. 


			Agent 				Environment


			Percepts      <--- Sensors      <----
		        |
Internal State		|
			|
World model 	        |
			|
Action knowledge	|
			|
			V
		       Current state 
		       of the world
                        |
                        |
condition-              |
action rules            |
                        |
                        V
                       Actions		---> Effectors


Advantages: 
uses history and other knowledge to circumvent some problems in 
worlds which may not be entirely perceptually accessible 

Disadvantages: 
requires larger memory

%----------------------------------------------------------------------------

Agent Architecture 2: Goal-Driven Agent

Goal-based agents use goals to select among many possible actions.

			Agent 				Environment


			Percepts      <--- Sensors      <----
		        |
Internal State		|
			|
World knowledge	        |
			|
Action knowledge	|
			|
			V
		       Current State
		       of the world
                        |
World knowledge	        |
                        |
Action knowledge        |
                        |
                        V
                       Future State 		
		       of the world 
                        |
Goals                   |
                        |
                        |
                        V
                       Actions		---> Effectors


Advantages: rational agent that choose actions based on its goals

Disadvantages: design is complex; requires more processing

%----------------------------------------------------------------------------

The question is not which is the correct design for intelligent
agents.  The issue, instead, is what design is appropriate for kind of
world, for kind of sensors and effectors, and for what kind of
computational constraint (e.g. limited processing time)

For example, an artificial frog that catches artificial flies may well
be based on the first design above because the world is perceptually
accessible, but there is great need for rapid processing.

%----------------------------------------------------------------------------

Copyright (c) Ashok Goel 1997 
College of Computing 
Georgia Institute of Technology, 
Atlanta, Georgia 30332-0280 
goel@cc.gatech.edu


%----------------------------------------------------------------------------