Transcript IofT 1910 W Fall 2006 Week 4

IofT 1910 W Fall 2006
Week 5
Plan for today:
 discuss questions asked in writeup
 talk about approaches to building
intelligence
 talk about the lab on Thursday
1. Fundamental principles used in
the paper
2. Control architecture
 Model of intelligence
 Traditional model
 Brook’s model
 Model used in this paper
 Control architecture
 Subsumption
 Finite State Machine
Model of intelligence
Traditional model, according to Brooks
 In the traditional model
cognition mediates
between perception and
actions.
 Actions affect the world
 Perception gets
information from the
world to feed cognition
Perception
World
Cognition
Action
Brooks’ model
 Cognition is in the
eye of the
observer.
 Perception and
action do all the
work.
 The world is its
own best model.
World
Action
Perception
Cognition
Model in today’s paper?
AI = Rational Agents
An agent is an entity that perceives and acts.
More abstractly the agent is a function from percept
histories to actions.
The material on Agent architectures is from
Russell&Norvig,”Artificial Intelligence: a Modern Approach”, Prentice-Hall, 2003
Agent types
 We can design different agent types, from
simple to more complex.
 Agents operate in different environments
(observable vs partially observable,
deterministic vs stochastic, static vs
dynamic, episodic vs sequential, discrete vs
continuous, single agent vs multi-agent,
etc).
 The environment mostly determines the
agent design.
The vacuum cleaner world
The vacuum cleaner can perceive location and contents
(in A, dirt).
Actions it can do are: left, right, suck, no-op.
How should the agent decide what to do?
Agent types: A simple reflex agent
The agent reacts to the environment using its rules,
but has no memory of its past actions.
Agent types: A reflex agent with
state
The agent reacts to the environment using its rules,
and has memory of its past actions.
Agent types: A learning agent
The agent can adapt its actions to increase performance.
Robot control architecture

At least two approaches:
1. The subsumption architecture, where behaviors are built by
successive layers of modules, each of which is a Finite State
Machine (FSM). A Subsumption Architecture builds a system
by layering levels of control, allowing lower levels to override
the higher ones and injecting higher level outputs into lower
levels.
2. A finite state machine where states are connected by state
transition links and where each state includes multiple
behaviors. States allow decomposition of complex systems
into small chunks. Transitions handle flow control.
 Either one can be used. People use state machines more
often than the subsumption architecture, because they are
more flexible.
FSM vs. Subsumption
FSM
(ExploreMachine)
Brooks’ Subsumption Control System
walk
No Obstacle
Obstacle
Detected
turn
Slide from “Robotics Seminar CSI445/660, Spring 2005
Robert Salkin & Shawn Turner, University at Albany, SUNY
An example of an Agent using FSM:
MinDART
TAKE TARGET
HOME
SEARCH FOR
TARGETS
Obstacle Avoidance
Target
Dropped
Purposeful Search
Detect Other Targets
Obstacle Avoidance
Random Search
Homing
Target
Found
Obstacle Avoidance
Target Alignment
Pick Up Target
GRAB TARGET
Target
Grabbed
Lab on Thursday
 Where: in EE/CS 2-140
 You need your U card to get in.
 We’ll start learning URBI, a scripting
language that can be used either with
a memory stick or with the wireless
network.
 You can also continue using MEdit and
learn how to add sounds and to
program the LEDs.