Transcript Introduction to Artificial Intellilghee 2 Intelligence Agents.

Introduction to Artificial Intelligence
2 Intelligence Agents.
John Woodward
[email protected]
http://www.cs.nott.ac.uk/~jrw/
Agents, Environment,
Actuators.
•
•
•
•
An agent perceives its
environment and
manipulates it using actuators.
E.g. a human has sensor organs (eyes and ears
…) and actuators (hands, legs, …).
• E.g. a robot has cameras and range finders for
sensors and wheels and graspers for actuators.
• Software agents on the internet…
Percept and Perceptual Sequence.
• Percept is the agents inputs at any one instance in
time.
• Perceptual sequence is the complete history of
percepts (i.e. a list of input data).
• An agent’s actions are based on perceptual
sequence.
• An agent function maps the perceptual sequence to
an action.
• We could do this as table, but we implement an
agent program, which implements the function.
Vacuum-Cleaner
•
•
•
•
•
Vacuum agent perceives
which square (A or B)
and if clean or dirty.
It has actions; move left/right, suck, do nothing.
One simple function; if current square dirty, then
suck, else move to other square.
• We can write perceived state and action pairs
• [A, Clean] right (if in A && clean, then move right)
• [A, Dirty] suck (if in A && dirty, then suck)
Performance Measure.
• A performance measure evaluates a sequence of
environmental states.
• It is not easy to design a suitable measure.
• It is better to design performance measures according
to the what is actually wanted in the environment,
rather than how you think the agents should behave.
• E.g. rewarding robot for picking up rubbish, will just
empty the bins and pick up the rubbish again, rather
than doing what you want and cleaning.
Rational Agent
Rational behavior depends on
1. the performance measure.
2. The agent’s prior knowledge of the environment
3. The actions the agent can perform
4. The agent’s percept sequence.
An agent acts rationally if it uses (performs the action)
what is knows (prior knowledge and perceptual
sequence) to give the maximum performance measures
Vacuum-Cleaner World
1.Performance e.g. one point for each clean
square over 1000 time steps.
2. The environment is know a priori (e.g. only two
squares, we cannot move outside, after we
clean then the square is clean and not dirty).
3. Only actions are LEFT, RIGHT, SUCK.
4. The agent correctly perceives its location and if
that square is clean or dirty.
How else can this vary?
Dung Beetle & Sphex Wasp
• A dung beetle plugs its nest with a
ball of dung. If the ball of dung is
removed from the beetle as it is
carrying it – it still continues to
the nest without changing its
actions.
• A wasp drags a caterpillar to its
nest, but check the hole first. If
the caterpillar is removed, the
wasp starts the process again and
recheck the hole
Autonomy.
• A rational agent should be
autonomous.
• It should learn, to compensate for
partial information or information it
lacks (exploration vs. exploitation).
• Nature “hard codes” some information
in individuals (nature verse nurture).
• E.g. feral children, universal grammar,
new born horse. Independent offspring.
• Incorporation of learning allows for the
design of a single agent which will
perform well in a variety of
environments.
2.3 Nature of Environments.
2.3.1 specifying the task environment.
• Performance: what are we measuring
performance against?
• Environment: what is the environment the
agent will operate in?
• Actuators: how can the agent affect the
environment?
• Sensors: how does the agent get information
from the environment?
2.3.2 Properties of task environment
• Fully observable vs. partially observable.
• Single agent vs. multiple agent. (competitive
vs. cooperative)
• Deterministic vs. stochastic.
• Episodic vs. sequential.
• Discrete vs. continuous.
• Known vs. unknown.
2.4 Structure of Agents
• AI is largely about designing an agent
program which controls the agent.
Agent = architecture + program
Trigonometric tables (lookup tables) can
be compressed into a short algorithm.
Intelligence is not a look up table (we
could not list all of the English
sentences – just too many).
Can AI write programs that produce
intelligent behavior?
2.4.2 simple reflex
agents
• Simplest agent
• Action is based on
Current percept only.
IF condition THEH action
E.G. IF car in front brakes THEN we brake.
Human reflexes e.g. blinking, knee jerk.
Will not be optimal if the environment
is partially observable.
Random actions can escape from possible infinite
loops (e.g. a robot stuck between two walls),
2.4.3 model-based
reflex agents
Keeping a model of the world
as an internal state avoids
problems with partially observable worlds.
• To do this we need to know two things
1. how the world changes independently of the
agent (e.g. night follows day).
2. how the agent’s actions affect the world (e.g.
switching on a light will make a room bright).
2.4.4 goal-based
• There is a desirable
goal state of the
world.
• Sometime it is easy to
describe the sequence
of actions to achieve a
goal, but sometimes
more difficult e.g. a
Rubik cube
2.4.5 utility based
agents
• Goal states just provide
Yes/no binary information.
A more general version
Is a utility function.
A utility function is essentially an internalization of
the performance measure. It helps in two ways;
1. A goal state may not be achievable with certainly
2. There may be multiple conflicting goals.
A rational agent maximizes expected utility.
2.4.6 general
learning agent.
• Instead of writing the
components ourselves,
why not let the machine
learn? Turing (1950).
The learning element changes the way the
agent performs. The critic affects
learning. And a problem generator
suggest new temporarily possible
suboptimal actions which may perform
better eventually.
Pacman
The winners of the best 2nd year group project prize have had a paper accepted at an
IEEE conference. It discusses the methods they used to develop an automated Pacman
player.
The paper, Ghost Direction Detection and other Innovations for Ms. Pac-Man,
has been accepted for the 2010 IEEE Conference on Computational Intelligence and
Games, 18-21 Aug 2010, Copenhagen, Denmark. The authors are Nathaniel Bell,
Xinghong Fang, Rory Hughes, Graham Kendall, Edward O’Reilly and Shenghui Qiu.
The paper will eventually appear in IEEE Xplore.
One of the students (Xinghong Fang), will attend the conference to present the paper
and also to enter the automated Pacman player competition, which the software was
developed to address.