June 2004

        

AI Expert Newsletter

June 2004

Dennis Merritt

AI - The art and science of making computers do interesting things that are not in their nature.

Software Downloads

Some of the software described in these newsletters is now available for download at www.ainewsletter.com/downloads.

Fantasy Games

Adventure was the first computer fantasy game. It had all of the elements we see today in Nintendo and Play Station fantasy games, where the computer has a model of an environment and the player manipulates a character in the game, travelling from place to place, encountering various challenges and puzzles along the way, collecting points for challenges met, and in general being entertained by unfolding story line and discovered scenarios.

The only difference was that instead of graphics, Adventure had textual descriptions of the places and situations, and instead of a mouse or controller, the game was driven with simple English.

Adventure captivated the computing community of the time. Despite being hand distributed on a bulky magnetic tape requiring system programmer installation, it spread like a modern virus and programmers around the World were soon trying to figure out how to slay the dragon and get the plover's egg, instead of whatever programmer's were supposed to be doing.

Adventure inspired others to start to write other fantasy games, with Zork probably being the most famous successor. The creators of Zork eventually started InfoCom, a company whose business was the creation of these games in a genre they named interactive fiction (IF).

As graphics became more accessible, the text-based interactive fiction gave way to the more modern fantasy games, but the ideas were still basically the same, adventurers travelling about, solving puzzles, etc.

In researching this newsletter, I was pleased to discover that the IF community is still alive and well. The tools for creating IF are easily accessible to all and are being applied in the creation of a growing library of clever, well-written, and entertaining games. The links provide a number of sites describing the past, present, and future of IF, as well as a wealth of great games to play.

Game Development Tools

Adventure was, in many ways, an early AI program, in that it interacted with the player in an intelligent manner, understood simple English, changed the game enviroment based on those commands, and remembered the state of the game. So once a player moved to a room, or moved items in a room, or broke something, those changes would be reflected in the dialog and descriptions the game used to communicate with the player.

The game fit this newsletter's definition of AI, in that it was an interesting application that was difficult to write because it did not fit cleanly into a computer's nature. The development of IF provides a classic insight into the way AI tools in general are created.

The first game was written in FORTRAN (see the code corner for details). But pure FORTRAN code, like most procedural programming languages, didn't have the expressive capability necessary to describe the environment. So the developers had to create something like a virtual machine from a combination of data structures and FORTRAN code.

The data structures were used as a declarative programming language for the game. They made it easy to add new rooms, objects, and conditions on which travel between them could be made. They also made it easy to create all of the text describing the places, things and situations.

But they couldn't handle everything, so if-then-else statements were still used to handle the most out-of-the-ordinary situations in the game.

So the original code really had three components:

  • declarative data structures which were a primitive language for coding most of the game;
  • an interpreter that used those data structures for driving the game;
  • interspersed rules in the form of if-then-elses that added variety and interest.

When InfoCom got into the serious business of creating IF, the first thing they did was design better tools. They created a language purely for writing IF. And, of course, it required an interpreter to run the game.

In other words, like for many AI disciplines, they created a virtual machine with its own programming language for the purpose of building a certain type of intelligent application. The machine is called a Z-machine, and is available from a number of sites and can be used to run the numerous games built for it.

As the market for interactive fiction shrunk, the technology moved to the hobbyists, and a number of tools have since evolved for the creation of IF games. Inform, Hugo, and TADS are probably the big three and you can find information about them in the links.

There also appear to be numerous tools available for the creation of graphic fantasy games, but I didn't look very closely at them for this issue. This is in part because of a fascination with the media of interactive fiction. Just as a book presents a story in a very different way from a movie, so does a text-based fantasy game present an adventure in a very different way from a graphic one.

While the graphics in a game like Myst are fantastic, they don't create the bond between author and player that IF does. In IF the game itself talks to you, and has a personality. You wind up with a sense of trying to match wits with the game writer that you simply don't get from graphics, much as a movie can't reveal a character's personality as well as a book can.

And the writers of IF today are working with better tools and bringing excellent writing and puzzle creation skills to the genre. Here's a sample of some of the writing from Max Blaster and Doris de Lightning Against the Parrot Creatures of Venus, by Dan Shiovitz and Emily Short:

Secret Meeting Chamber (sitting on your chair)

Deep beneath the streets of Astro-City, there lies a maze of deadly traps and
deadlier guards, inside of which is a secret, hidden fortress, inside of which
is a dark and mysterious forest (albeit a rather small one which needs to be
lit by artificial lights and tended by a gardening crew once a week) that conceals
a tiny cottage, inside of which are a dozen angry leopards[2]. It is because
of this careful attention to detail, no doubt, that nobody has ever discovered
the trapdoor beneath the cottage that leads to this meeting-room. But it is
here that you and your fellow scientists sit around the great oak table in the
center of the room and plot your schemes and scheme your plots.

And later as Max meets Doris:

You beam down at Doris, then transfer your gaze to the Director. “So who's
the little lady?”

There's a pensive silence while Doris grinds her teeth and the Director taps
his cigar. The vacuumbot whirrs sadly and drifts towards the floor to collect
the ashes. “Max,” says the Director heartily, “meet Doris de
Lightning. Your new partner.”

This game is particularly noteworthy, and won awards, for its NPCs or non-player-characters. The player can either choose to work the game as Max or Doris, and either way, the other character becomes very much a part of the game.

Emily Short's Web site (see links) contains a number of insightful articles on the various technology challenges in creating IF, including one analyzing NPCs, the various approaches authors can use, and the tools available for creating believable NPCs. NPCs were a big part in the first Adventure as well, in the form of dwarves and the pirate.

Code Corner

Adventure in FORTRAN

This newsletter defines AI as being the art and science of making computers do interesting things that are not in their nature. As described in the first newsletter, this usually involves the creation of a virtual machine with a nature better suited to the application.

AI programming, then, would involve the creation of a virtual machine and then programming that machine to build the application. So the question of what tool is best for AI really refers to the tools for building the virtual machine.

For building the virtual machine, it really doesn't matter what tools are used. Any programming langauge will do. In the case of the original Adventure, the language of choice was FORTRAN. The FORTRAN code was higher level code, that dealt with most of the generalities of moving about the environment.

Here are the coded messages for various scenarios:

32 YOU CAN'T GET BY THE SNAKE.
33 YOU ARE IN A LARGE ROOM, WITH A PASSAGE TO THE SOUTH, A PASSAGE TO THE
33 WEST, AND A WALL OF BROKEN ROCK TO THE EAST. THERE IS A LARGE "Y2"    ON
33 A ROCK IN THE ROOM'S CENTER.
34 YOU ARE IN A JUMBLE OF ROCK, WITH CRACKS EVERYWHERE.
35 YOU'RE AT A LOW WINDOW OVERLOOKING A HUGE PIT, WHICH EXTENDS UP OUT OF
35 SIGHT. A FLOOR IS INDISTINCTLY VISIBLE OVER 50 FEET BELOW. TRACES OF
35 WHITE MIST COVER THE FLOOR OF THE PIT, BECOMING THICKER TO THE RIGHT.
35 MARKS IN THE DUST AROUND THE WINDOW WOULD SEEM TO INDICATE THAT
35 SOMEONE HAS BEEN HERE RECENTLY. DIRECTLY ACROSS THE PIT FROM YOU AND
35 25 FEET AWAY THERE IS A SIMILAR WINDOW LOOKING INTO A LIGHTED ROOM. A
35 SHADOWY FIGURE CAN BE SEEN THERE PEERING BACK AT YOU.
36 YOU ARE IN A DIRTY BROKEN PASSAGE. TO THE EAST IS A CRAWL. TO THE
36 WEST IS A LARGE PASSAGE. ABOVE YOU IS A HOLE TO ANOTHER PASSAGE.
37 YOU ARE ON THE BRINK OF A SMALL CLEAN CLIMBABLE PIT. A CRAWL LEADS
37 WEST.
38 YOU ARE IN THE BOTTOM OF A SMALL PIT WITH A LITTLE STREAM, WHICH
38 ENTERS AND EXITS THROUGH TINY SLITS.
39 YOU ARE IN A LARGE ROOM FULL OF DUSTY ROCKS. THERE IS A BIG HOLE IN
39 THE FLOOR. THERE ARE CRACKS EVERYWHERE, AND A PASSAGE LEADING EAST.
40 YOU HAVE CRAWLED THROUGH A VERY LOW WIDE PASSAGE PARALLEL TO AND NORTH
40 OF THE HALL OF MISTS.
41 YOU ARE AT THE WEST END OF HALL OF MISTS. A LOW WIDE CRAWL CONTINUES
41 WEST AND ANOTHER GOES NORTH. TO THE SOUTH IS A LITTLE PASSAGE 6 FEET
41 OFF THE FLOOR.
42 YOU ARE IN A MAZE OF TWISTY LITTLE PASSAGES, ALL ALIKE.
43 YOU ARE IN A MAZE OF TWISTY LITTLE PASSAGES, ALL ALIKE.
44 YOU ARE IN A MAZE OF TWISTY LITTLE PASSAGES, ALL ALIKE.
45 YOU ARE IN A MAZE OF TWISTY LITTLE PASSAGES, ALL ALIKE.
46 DEAD END
47 DEAD END
48 DEAD END
 

Here's the messages for various objects that depend on the state of the game:

8 LITTLE BIRD IN CAGE
000 A CHEERFUL LITTLE BIRD IS SITTING HERE SINGING.
100 THERE IS A LITTLE BIRD IN THE CAGE.
9 *RUSTY DOOR
000 THE WAY NORTH IS BARRED BY A MASSIVE, RUSTY, IRON DOOR.
100 THE WAY NORTH LEADS THROUGH A MASSIVE, RUSTY, IRON DOOR.
10 VELVET PILLOW
000 A SMALL VELVET PILLOW LIES ON THE FLOOR.
11 *SNAKE
000 A HUGE GREEN FIERCE SNAKE BARS THE WAY!
100 >$< (CHASED AWAY)
12 *FISSURE
000 >$<
100 A CRYSTAL BRIDGE NOW SPANS THE FISSURE.
200 THE CRYSTAL BRIDGE HAS VANISHED!
 

Here's some code that defines the rooms, connections, and conditions under which some of the connectins can be used:

3
1 2 2 44 29
1 3 3 12 19 43
1 4 5 13 14 46 30
1 5 6 45 43
1 8 63
2 1 2 12 7 43 45 30
2 5 6 45 46
3 1 3 11 32 44
3 11 62
3 33 65
3 79 5 14
 

Here's various responses to various commands based on situation:

163 THE BEAR LUMBERS TOWARD THE TROLL, WHO LETS OUT A STARTLED SHRIEK AND
163 SCURRIES AWAY. THE BEAR SOON GIVES UP THE PURSUIT AND WANDERS BACK.
164 THE AXE MISSES AND LANDS NEAR THE BEAR WHERE YOU CAN'T GET AT IT.
165 WITH WHAT? YOUR BARE HANDS? AGAINST *HIS* BEAR HANDS??
166 THE BEAR IS CONFUSED; HE ONLY WANTS TO BE YOUR FRIEND.
167 FOR CRYING OUT LOUD, THE POOR THING IS ALREADY DEAD!
168 THE BEAR EAGERLY WOLFS DOWN YOUR FOOD, AFTER WHICH HE SEEMS TO CALM
168 DOWN CONSIDERABLY AND EVEN BECOMES RATHER FRIENDLY.
 

And here's some FORTRAN code to deal with the special situation of dwarves, pirates and the troll bridge:

C TRAVEL 303. TROLL BRIDGE. MUST BE DONE ONLY AS SPECIAL MOTION SO THAT
C DWARVES WON'T WANDER ACROSS AND ENCOUNTER THE BEAR. (THEY WON'T FOLLOW THE
C PLAYER THERE BECAUSE THAT REGION IS FORBIDDEN TO THE PIRATE.) IF
C PROP(TROLL)=1, HE'S CROSSED SINCE PAYING, SO STEP OUT AND BLOCK HIM.
C (STANDARD TRAVEL ENTRIES CHECK FOR PROP(TROLL)=0.) SPECIAL STUFF FOR BEAR.

30300 IF(PROP(TROLL).NE.1)GOTO 30310
   CALL PSPEAK(TROLL,1)
   PROP(TROLL)=0
   CALL MOVE(TROLL2,0)
   CALL MOVE(TROLL2+100,0)
   CALL MOVE(TROLL,PLAC(TROLL))
   CALL MOVE(TROLL+100,FIXD(TROLL))
   CALL JUGGLE(CHASM)
   NEWLOC=LOC
   GOTO 2
 

Adventure in Inform

Inform is one of the languages used to create today's IF. It is an object-oriented language that lets an author use a C-like syntax for creating all of the objects in a game, and the messages associated with them and the actions that can be applied to them.

The language contains ready-made support of many of the concepts useful in IF, such as whether a room is lit or not, and the kinds of things that are visible in containers of things.

Here's a bite-sized sample from "The Last Sonnet of Marie Antoinette" by Emily Short.

Mobile nutsack "small sack" woodbox,
    with name "sack" "small" "cloth" "bag",
    mater CLOTH,
    size 1,
    shape LINEAR,
    description [;
        "A small cloth bag drawn shut with a drawstring."; ],
    has container openable ~open;

Mobile nut "nut" nutsack,
    with name "nut",
    mater FOOD,
    size 1,
    shape ROUND,
    description [;
        "A single round nut."; 	],
    after [;
        eat:"Hmm.  It's still hard to say exactly what kind of nut that was.";]; 

The full source is available from Emily Short's Web site, and illustrates a lot of the expressive power of Inform. In this particular game a class called 'Cheese' is created with all sorts of behaviors associated with cheese, and then a number of subclasses of different types of cheese, such as brie and parmesan are created as well. I haven't played the game, but there is all sorts of wonderful text describing the various cheeses, plus what look like complex behaviors resulting from eating the various types.

Adventure in Prolog

It was my early experience with Adventure that led to my current involvement with AI. I was in the process of designing my own Inform like tools using C that I happened on a lecture on Prolog. Given that a large portion of a game involves keeping the state of the game and pattern-matching and search algorithms, it seemed as if Prolog would be a tremendous productivity aid in writing IF.

I learned Prolog explicitly to create IF, and quickly discovered two things:

  • Prolog was an excellent tool for the job; and
  • I lacked the imagination necessary for creating fiction.

So instead I created a game that modelled a simple subset of my real life, and wrote a book on Prolog using it as an example, which lead to other Prolog applications and a career in AI in general.

As in so many aspects of life, the media shapes the message. Today's IF writers have excellent tools like Inform, but they shape the resulting games by the features and manner of expression they provide.

Likewise, creating a game in Prolog provides a different set of tools for the writer. As a base language, Prolog doesn't have all the built-in support for standard IF features, as Inform does. On the other hand, native Prolog syntax adapts much more readily to coding IF than say FORTRAN or C++.

Here is one way to represent the rooms and objects in a simple game in Prolog:

% Initial facts describing the world. Rooms and doors do not change,
% so they are compiled.
room(office).
room(kitchen).
room('dining room').
room(hall).
room(cellar).
door(office,hall).
door(hall,'dining room').
door('dining room',kitchen).
door(kitchen,cellar).
door(kitchen,office).
connect(X,Y):-
   door(X,Y).
   connect(X,Y):-
   door(Y,X).
% These facts are all subject to change during the game, so rather
% than being compiled, they are "asserted" to the interpreter at
% run time. This predicate is called when "nanisrch" starts up.
init_dynamic_facts:-
   assertz( location(desk,office) ),
   assertz( location(apple,kitchen) ),
   assertz( location(flashlight,desk) ),
...
   assertz( location(crackers,desk) ),
   assertz( location(broccoli,kitchen) ),
   assertz( here(kitchen) ),
   assertz( turned_off(flashlight) ).
furniture(desk).
furniture('washing machine').
furniture(table).
edible(apple).
edible(crackers).
tastes_yuchy(broccoli).
 

And here is the rule that implements the verb 'eat'. First, you have to 'have' the thing you want to eat, and then there are three cases for things that are edible, taste yuchy or are something else. This leads to attempts at humor when the three-year old protagonist tries to eat broccoli.

% eat, because every adventure game lets you eat stuff.
eat(Thing):-
   have(Thing),
   eat2(Thing).
eat(Thing):-
   respond(['You don''t have the ',Thing]).
   
eat2(Thing):-
   edible(Thing),
   retract(have(Thing)),
   respond(['That ',Thing,' was good']).
eat2(Thing):-
   tastes_yuchy(Thing),
   respond(['Three year olds don''t eat ',Thing]).
eat2(Thing):-
   respond(['You can''t eat a ',Thing]).
 

Because Prolog is a general purpose language, it can be used to model any sort of environment and interaction the author desires. For example, one of the articles reachable from the links describes a feature not available in current IF authoring tools, which is a sense of physical space. In other words, the connections between places are defined, but not absolute x-y-z coordinates.

If the places had x-y-z coordinates, then a game could utilize tunneling as a general feature. Now tunneling could be simulated in current tools by simply having a closed connection which is opened using a 'tunnel' verb, but if one were writing in a more general purpose language, one could create a model of places that included coordinates as well as connections, and let the player discover that maybe tunneling up from the cellar will lead to the vault with the safe.

Other types of AI technology could be used to model various aspects of the game as well. Maybe fuzzy logic controllers could play a role? Or Bayesian probability? How about neural nets in a non-playing-character that can be trained during the course of the game? So a player could have a dog that learned?

The other area where it appears, and I say appears because I really haven't examined Inform and the others in much depth, general AI techniques might help is in the game-player interaction.

It seems like some of the chatbot technology discussed in other newsletters might provide good interfaces for IF, especially in conjunction with an ontology that described the objects in the game. Another option is DCG (definite clause grammar) to provide better natural language parsing, as used in the Adventure in Prolog (see links) tutorial. The point being that these different sorts of language technologies could easily be experimented with in the context of a particular game.

Here's an example of some DCG used to parse the player's commands that uses the current context of the game to interpret the commands. In particular the ambiguous use of the word 'light' can be interpreted to mean a room light or the flashlight, depending on whether the player has the flashlight or not.

% Nouns are defined as rooms, or things located somewhere. We define
% special cases for those things represented in Nani Search by two
% words. We can't expect the user to type the name in quotes.
noun(go_place,R) --> [R], {room(R)}.
noun(go_place,'dining room') --> [dining,room].
noun(thing,T) --> [T], {location(T,_)}.
noun(thing,T) --> [T], {have(T)}.
noun(thing,flashlight) --> [flash,light].
noun(thing,'washing machine') --> [washing,machine].
noun(thing,'dirty clothes') --> [dirty,clothes].
% If the player has just typed light, it can be interpreted three ways.
% If a room name is before it, it must be a room light. If the
% player has the flash light, assume it means the flash light. Otherwise
% assume it is the room light.
noun(thing,light) --> [X,light], {room(X)}.
noun(thing,flashlight) --> [light], {have(flashlight)}.
noun(thing,light) --> [light].
 

The Prolog approach described here is not unlike the FORTRAN approach described above. The game is implemented using a combination of:

  • Prolog data structures,
  • An interpreter for those structures, and
  • Rules encoded directly using Prolog's native rule syntax.

If one were serious about using Prolog as a tool, one would quickly evolve towards using frame-like structures, as done for other AI tools described in previous newsletters, to model a complete language of game specification.

User Interface in IF

The biggest criticism of IF is probably the frustration caused by using what appears to be natural language. This is a problem in general with natural language.

Any person-computer interaction requires a cultural adjustment on either the computer's or the person's side. In other words, either the computer has to learn to understand 'person' or the person needs to understand 'computer'.

A natural language interface to a program is an attempt to program a computer to understand 'person'. Once having parsed and understood a human sentence, the computer then maps that sentence into one of a finite set of commands the particular application is prepared to carry out.

Menus are the exact opposite. The computer displays exactly all the things it can do, and the person then picks one from the menu. It is the person's responsibility to understand what the computer can do and the right menu items to choose.

Given that people can learn to understand computers easier than computers can learn to understand people, menus are by far the most popular human-computer interfaces.

I think IF would benefit from adapting this sort of approach, where the possible actions and objects are presented in menus. The possible actions would only be available once the player had discovered the right aspects of the game. So, for example, in Doris' adventures on Venus the fact that she can fly over the canyon using her rocket pants would only appear on the menu after the player had noticed she was wearing rocket pants.

Another useful feature would be the automatic drawing of a map as the player explored more of the game. And maybe using HTML links in the text to let the player more easily examine the items presented in the text describing a place. Apparently there are many many extensions to Inform, and probably Hugo and TADS as well, that provide these types of additional features for the gaming author.

Links

http://www.rickadams.org/adventure/ - Rick Adams wonderful site full of information about the origins of the first computer fantasy game. It includes downloads of various versions of the game.

http://www.ifiction.org/ - Dave Walton's site for interactive fiction. It has a version of Infocom's Z-machine designed for the Web, with nice features such as keeping the state of the game on the Web page so you can simply save the page when you want to stop and restart later. It has hundreds of interactive fiction games including the early Zork classics and Scott Adam's, of the Hitchhiker's Guide to the Universe fame, work.

http://www.csd.uwo.ca/Infocom/ - Peter Scheyen's site dedicated to the interactive fiction games of Infocom.

http://www.inform-fiction.org/ - Inform is a design system for creating interactive fiction. The Web site is full of resources, including tutorials on using Inform and other similar tools; references to other tools; and places to get Z-Machine interpreters for running games.

http://emshort.home.mindspring.com/ - Emily Short's, one of the best modern creators of interactive fiction, Web site.

http://teladesign.com/tads/ - The home page for the Text Adventure Development System (TADS) a sophisticated C-like tool for developing interactive fiction.

http://www.tads.org/ - The original home page for TADS.

http://www.amzi.com/AdventureInProlog/index.htm - The Prolog tutorial I wrote based on my experiences playing with IF using Prolog. The preface has some background on my early experiences with Adventure.


As always, feedback ideas and especially interesting links are welcome. Past issues are available at either www.ddj.com or www.ainewsletter.com.

Until next month,

Dennis Merritt


Copyright ©2002-04 Amzi! inc. and CMP. All Rights Reserved.