Notes for August. 22, 2001
Office hours:
CCB 119 MW: 1:30 to 3:00 PM
CRB 394 TR: 3:00 to 4:30PM
Student answers to "What is an Agent?"
- Someone/thing acting on behalf of someone/thing else.
- Someone/thing that assists to do something.
- Acts on minimal instructions
Fundamental faculties of intelligence:
- acting
- ex: going to a chair and picking something up; intelligence results in
some action
- sensing
- understanding, reasoning, learning
In order to act you must sense; blind action is not a characterization of
intelligence. So there is a connection b/w acting and sensing - this is
robotics, but robotics itself doesn't need understanding.
robotics: ability to sense and change the world.
Sensing needs understanding to be useful. ex: see professor and recognize
that he teaches this class.
perception: understanding what you sense.
AI: understanding without worrying about sensing and acting.
Agent: must sense and must act.
Intelligent Agent: must sense, must act, must be autonomous(to some
extent), must be rational.
simple intelligent agent ex: thermostat, auto light-turn on when someone
enters the room.
sense: temp, person enter
act: turn on heater, turn on light
These are agents, but they don't have memory, learning, etc.
Percept-based agent:
- information comes from sensors - percepts
- changes the agents current state of the world
- triggers actions through the effectors
this type of agent has become popular and has different names:
cognitive science - situated action
robotics - reactive control
but it has no notion of history; the current state is as the sensors see
it right now.
criteria for deciding how good a design is:
- tasks - ex: make a cup of coffee
- environment - what if you are out camping without any coffee
beans...need to grow the beans!
- efficiency
- processing - quickly make a decision. ex: frog eating a fly...good
design because sensors match input well
- execution - quickly achieve its goal.
- cost
- match between sensors and environment - can't use sonar in soundless
area (space)
Percept-based agents are not good for multiple, opposing,
goals.
ex: want to throw away a box on a chair
goal1: robot needs to pick up the box.
goal2: robot needs to avoid hitting the chair.
the robot might get stuck going forward and back because the goals are
opposing - this is:
Local Minima: getting stuck without achieving the main goal.
Problems with simple agent:
- Sub-optimal (non-efficient)
- Get stuck in local minima
State-based agent:
- information comes from sensors - percepts
- changes the agents current state of the world
- based on state of the world and knowledge (memory), it
triggers actions through the effectors
advantages:
- help avoid previous mistakes
disadvantage:
- processing is not as efficient
- knowledge, if uncertain or incomplete, can lead to errors.
- the more there is in memory, the harder it is to learn.
- we need a theory of learning, theory of memory.
Goal-based agent:
- information comes from sensors - percepts
- changes the agents current state of the world
- based on state of the world and knowledge (memory) and
goals/intentions, it chooses actions and does them through
the effectors.
goals demand a certain set of actions as opposed to others.
knowledge gives you a better state of the world.
goals help us choose better actions.
Q: Aren't knowledge and goals the same things as percepts and
action? They are both coded the same way, aren't they?
A:simple agents: simple knowledge/goals
complex agents: complex knowledge/goals
we will program it the same way, but in state/goal-based it will be more
complex.
Rationality: goal-directed behavior.
goal-based:
pros:more optimal solutions. higher efficiency of execution. - for most
tasks...
cons: more processing. when is it worth the extra processing (ex: lion
about to eat you and you think about what to do)? how does it learn
goals?
The progression of AI is not the way we described it.
It started very complicated with memory and knowledge and everything in
it.
The early model:
perception -> cognition -> action.
in 80's and 90's they revised the model some. they said most of the time
it works like this:
perception -> action.
constant debate between these 2 views continuing even today.
Notes taken by Lushi (Yisrael Lowenstein)
