version: $Id: project3.txt,v 1.3 2006/11/12 19:48:36 ioerger Exp $

CPSC 625
Fall 2006 
Final Project

In this project, you will be implementing a knowledge-based decision-making
procedure for an agent in a simulation of the Wumpus World in Prolog.  You
will have to download and install your own local version of SWI Prolog.
(http://www.swi-prolog.org/).

Attached are two files: wumpus.pl, and agent.pl.  wumpus.pl defines how the
wumpus world works, computes the effects of actions, determines what senses
are perceived, and keeps track of the score.  You may not manipulate
wumpus.pl, nor use any of the internal state information.  agent.pl is a
starter file that defines a simple decision-making procedure based only on the
senses the agent perceives, which you should extend or modify with your own
rules.

Important functions wumpus.pl that you can call are: reset, show, and do.
reset generates a new random configuration of the world.  Show will give a
simple visual representation, and print out some info like what direction the
agent is facing and what it is holding.  However, this information is not
available to the agent; the agent only has access to the basic senses (like
stench, glitter, and bump).  calling do(Action) will change the state, retract
and assert new senses, and show the result.

In agent.pl, a simple driver is called go(N), where you can specify a certain
number of steps.  It calls decide(Action) to pick the most desirable action in
the current state and execute it (call do).  Right now, it just picks up gold
when it can, and otherwise does a random move (forward, or turn left or
right).  Your job is to flesh out decide so the agent makes more intelligent
decisions and gets the highest score possible.  You might to choose a
different stopping criterion for go, such as going until either the score 
has become >100 or =-1000, or stopping once both gold has been collected and
wumpus killed, as opposed to just going a finite number of steps.

Part of this task requires inferring (and remembering) the effects of actions.
The case of holding(gold) is given as an example.  There is a rule which
infers it when see glitter and the action is pickup(gold); if satisfied,
holding(gold) is asserted (and since there is no putdown action, it remains
true thereafter).  In agent.pl, and clock (counter) called now(X) is provided
called, which gets incremented with each action taken.  You can use this to
remember time-specific facts, like when a room was visited or when the wumpus
was killed.  The action(Act,Time) is an example of this, which is
automatically asserted by my_do, which is a wrapper around do.

Here are the specific rules of this version of the wumpus world.  There is 1
gun (U), 1 gun (G), 1 wumpus (W), and more than one pits (P).  The cave will
have dimensions MxN (don't assume 4x4), but the agent will initially be in a
random (!) location.  Therefore, you will have to rely on 'bump' for feedback,
and keep track of (relative) location on your own (perhaps as "offsets" from
initial coordinate like 0,0).  I'm not sure if this helps, but you can assume
the agent will be facing north to start with.  

The actions available are: forward, turn(left), turn(right), shoot,
pickup(gold), pickup(gun).  You have to be in the same room to pick something
up.  It will be carried (even both at the same time - there is no exclusion)
thereafter, and there is no putdown operation.  The gun only has 1 bullet, so
it can only be fired once, and to kill the wumpus, you have to be adjacent to
it and facing it.

The senses, which are the only feedback about the state you may use are:
glitter, breeze, stench, see_gun, bump, scream.  As in the book, stench and
breeze are felt in rooms adjacent to the wumpus or a pit, respectively.  You
only see glitter (or gun) if you are in the same room as the gold (gun).

You may not query the state to determine any other information, such as the
location of the agent, wumpus, gold, or pits.  Nor may you query what
direction you are facing, what you are holding, whether the wumpus is alive,
etc.  You have to keep track of all this yourself, based on the sense and your
actions.

Reset will NOT guarantee that all objects are initially in unique locations.
However, overlaps should be rare, and we'll consider them uninteresting
special cases.  Don't bother designing rules for them; you can even ignore
them if you like (just reset again).

The score(N) of the agent is updated as follows: +100 for picking up gold or
shooting the wumpus, -1000 for enter room with the wumpus or a cave, and -1
otherwise (time-penalty for each step).

See the attached transcript for an example. (Note, this is an example of
selecting the actions manually using 'do'.  Ultimately, you will want your
agent to run in an automated mode using 'go'.)