This is an example of a step by step improvement of the basic Reflex-Agent design.
Due to the bad performance of the pure reflex-agent descriped by Russel and Norvig in their book Artificial Intelligence in the Wumpus-World, two new agents were developed that perform far better in this harsh environmnet.
These agents memorize their last actions (steps, shots, etc.). Due to this memory the new agents can base their decissions on their previous experiences they made on a field - thereby they incorporate a very simple form of learning.
The two agents differ in the amount of memory cells and capabilties to process the previous experiences.
Therefore the agent can exclude the last move it made from the list of his next moves if it resulted in a bump percept. When the agent fires an arrow at the Wumpus it excludes the field it came from - which must be without a Wumpus - and any shot it just made from the directions it wants to fire at.
Beside the moves and shots the agent also memorizes the death-scream of the Wumpus - this way the agent always knows when it is save or not.
Therefore this agent can exclude all moves that resulted in a bump from his list of next moves. Another advantage is the possibility to shoot only in directions it did not try yet - thereby this agent will always kill the Wumpus with only three arrows/javelins. (The direction it came from is save -> leaving only three more field -> which are marked off one by one by a shot.)
The pure Reflex-Agent: Reflex1
The second Reflex-Agent - a knowledged Agent: Reflex2
The third Reflex-Agent - the wise Guy:
Reflex3
The ultimate Reflex-Agent: ...
Well?
Know any more possibilities to get more out of the
Reflex-Agent model ?
(And - yes - we know there are still possibilities )
Go on - program an agent - give it a try;
for information on programming such an agent see the
Setuppage - project
Documentation)
A Reflex Labyrinth with all these agents.
"Tell me - What does it show?"
"What's the point of all this?"
A Resume‰:
It is not always enough for a
Reflex-Agent only to act on the percepts it currently received.
In a complicated world - like the Wumpus-World - the pure reflex-agent will
find it difficult to survive or grab a single piece of gold.
Still even the most simple reflex-agent has a fair chance (to survive) -
but in most cases only if it freaks out at the very beginning
of the simulation and returns home
(via a "CLIMB")...
or in a simple environment - where live just isn't that hard.
Therefore an advanced agent should also memorize the percepts it received - along with the actions it executed to get these percepts (keep a history line). With this information an agent can far better dodge pits and elude the Wumps (if the agent has no arrows left to kill it).
The example also shows the difference
between live and death.
The simple reflex-agent might fall into a pit or be
gobbled by the Wumpus because of it's random
movement after the "BREEZE" and "STENCH"
percepts.
The newer agents can't fall into a pit/be gobbled by the Wumpus
because they can always take their last move back.
The simple reflex-agent is just lost - it doesn't memorize what it did
and therefore can't revert a move.
Last modified: Mo. June 09 14:07:05 1997 by Roman Valehrach
email to: webmaster@kr.tuwien.ac.at