This is an old revision of the document!


Modeling Games in GDL

The main goal of this class is to learn how to model turn games in a rule based language. After finishing the class, student should be able to model simple games and point out the difficulties related with this task.

1 Preliminaries

GDL (Game Description Language) is a language designed to describe turn games. It was created to be used in so called General Game Playing, a subdomain of the AI, focused on creating bots able to play in any game that can be formally described.

The class is based on a tutorial from Stanford and slides from the university in Dresden.

2 Modeling the World

There are two main approaches to model the world of the game:

  1. coarse-grained — where every state of the game is a state in a giant state machine. Every unique state corresponds to one state in the machine. The state machine may grow quite impressive. Beside that state doesn't have any structure so is no really helpful if you want to exploit it to solve the problem in an intelligent manner.
  2. fine-grained — state of the world is defined by so called fluents, which can be treated as variables. The state here is structured (the current values of the fluents) which we may use.

GDL belongs to the fine-grained family, so there are:

  • fluents (actions made by the players are a special kind of fluent
  • game rules, which are represented by… rules

3 Example: Tic-Tac-Toe

We will show the basic features of the GDL on the example of the popular tic-tac-toe game.

Players

Let's start with players “nought” and “cross”. There is a special fact role just to do that:

role(nought)
role(cross)

Syntax is the same as in the Prolog (or Datalog) language with the only difference, that there is no dot at the end of the sentence. All constants start with a lower-cased letter and variables are upper-cased.

Fluents

Now we should define fluents that make up the game state. In a Tic-Tac-Toe game there are 29 fleunts:

  • 1 fluent per every cell, telling if there is a cross (x) in the cell (9 fluents)
  • 1 fluent per every cell, telling if there is a nought (o) in the cell (9 fluents)
  • 1 fluent per every cell, telling if the cell is empty (blank - b) (9 fluents)
  • 1 fluent per player, telling if it is his turn to play (2 fluents).

To define the fluents one uses the base facts:



Here :- reprenst implication . & is a logical and. In natural language the first line means: “if M and N represent cells' indexes, then there is a fluent for cell with (M, N) coordinates, telling there is a cross in it”. If we didn't use rules, we could write all the fluents by hand:

base(cell(1,1,x))    base(cell(1,1,o))    base(cell(1,1,b))
base(cell(1,2,x))    base(cell(1,2,o))    base(cell(1,2,b))
base(cell(1,3,x))    base(cell(1,3,o))    base(cell(1,3,b))
base(cell(2,1,x))    base(cell(2,1,o))    base(cell(2,1,b))
base(cell(2,2,x))    base(cell(2,2,o))    base(cell(2,2,b))
base(cell(2,3,x))    base(cell(2,3,o))    base(cell(2,3,b))
base(cell(3,1,x))    base(cell(3,1,o))    base(cell(3,1,b))
base(cell(3,2,x))    base(cell(3,2,o))    base(cell(3,2,b))
base(cell(3,3,x))    base(cell(3,3,o))    base(cell(3,3,b))
base(control(nought))
base(control(cross))

Rules in GDL have a one special feature: they can recurse, but they can't loop forever. More details here.

Available Actions

Now, we should model available actions, there are 20 of them here:

  • 1 action per every cell, that the player puts the 'nought' in
  • 1 action per every cell, that the player puts the 'cross' in
  • 1 action noop per user, telling that player does do nothing

Thee noop action exists only in order to simulate the players' turns. When one player makes the action then the second one has to take the noop action.

In order to define an action we use the input predicate:

input(R, mark(M,N)) :- role(R) & index(M) & index(N)
input(R, noop) :- role(R)

The input predicates takes two arguments, first is the player, second is the action itself.

Knowing the available actions, we have to define when the action is a legal move with a legal predicate. We use the true predicate to check if the fluent is true at the moment..

legal(W,mark(X,Y)) :-
   true(cell(X,Y,b)) &
   true(control(W))

legal(nought,noop) :-
   true(control(cross))

legal(cross,noop) :-
   true(control(nought))

In other words - player may mark a cell only if it is empty and it is his turn. If it is not his turn, he can only perform the 'noop' action.

Initial State

Now we have to define the initial state of the game (draw the empty board). The init predicate serves this purpose. For the tic-tac-toe every cell should be empty, and the nought player should play first.

init(cell(M,N,b)) :- index(M) & index(N)
init(control(nought))

State Changes

Now the most difficult part — how does the state evolve in the game. We assume that the new state is created after the very move and depends only on the previous state and actions made by the players. On could notice the frame problem occurring here. The etymology of the frame is the movie tape, where every frame differs in details from the previous one. Similarly here, we would like the next state to be exactly like the previous one except some changes made explicitly by the players. The problem is that in our case we have also to model lack of changes :(

To define the state evolution we use the next predicate, defining the fluent state in the next turn. The 'does' predicate check what actions were made by the players during the current turn.

next(cell(M,N,x)) :-
  does(cross,mark(M,N)) &
  true(cell(M,N,b))
 
next(cell(M,N,o)) :-
  does(nought,mark(M,N)) &
  true(cell(M,N,b))
 
next(cell(M,N,W)) :-
  true(cell(M,N,W)) &
  distinct(W,b)
 
next(cell(M,N,b)) :-
  does(W,mark(J,K)) &
  true(cell(M,N,b)) &
  distinct(M,J)
 
next(cell(M,N,b)) :-
  does(W,mark(J,K))
  true(cell(M,N,b)) &
  distinct(N,K)
 
next(control(nought)) :-
  true(control(cross))
 
next(control(cross)) :-
  true(control(nought))

The defined actions mean that:

  1. if the cell is empty and the 'cross' player marked it then there will be a 'x' in this cell next turn
  2. if the cell is empty and the 'nought' player marked it then there will be an 'o' in this cell next turn
  3. if the cell wasn't empty (distinct checks equality of the arguments), then it will stay the same next turn
  4. if the cell was empty and wasn't mark this turn, it will be still empty next turn (two rules!)
  5. if the current turn belongs to the the 'nought' player, next will belong to the 'cross' player
  6. if the current turn belongs to the the 'cross' player, next will belong to the 'nought' player

It's worth to note, that actions 3 and 4 are designed only to cope with the frame problem.

Game Ending Conditions

Every player knows that there are two ways to finish the game: make a line or fill all the available cells. Now we will check if there was a line made by the player.



The true(line(naught)) query will tell if there is a line made of the 'o's on the board.

Now we can define the ending conditions with the terminal predicate:



Game ends when:

  • there is a line of crosses
  • there is a line of naughts
  • there is no empty cell

The '~' operator means negation. Similarly to the Prolog it follows the closed world assumption, ie. sentence is false if can't prove otherwise. In GDL ther are additional restrictions we want worry about for now (more info here).

Winning Conditions

Bots don't have feelings but we still should give them prize for the good game. And we do this with the goal predicate:

goal(cross,100) :- line(x) & ~line(o)
goal(cross,50) :- ~line(x) & ~line(o)
goal(cross,0) :- ~line(x) & line(o)
 
goal(nought,100) :- ~line(x) & line(o)
goal(nought,50) :- ~line(x) & ~line(o)
goal(nought,0) :- line(x) & ~line(o)

According to the rules above winner gets 100 (so much win!), in case of tie only 50, and loser gets 0. We don't what this number represent but it has to be something good like cookies.

There is no arithmetics in GLD — numbers are only used to define indexes, prizes, time, etc.

End

That's all folks, you've just modeled a simple tic-tac-toe game.

3.1 Assignment 1

Let's take an even simpler game:



Answer the following questions:

  1. how many players are there? (correct answer: 2, why?)
  2. how many fluents are defined? (correct answer: 4, why?)
  3. how many actions are there? (correct answer: 8, why?)
  4. how many legal moves does the 'white' player have in the initial state of the game? (correct answer: 3, why?)
  5. how many fluents are true in the initial state of the game? (correct answer: 1, why?)
  6. assume in the first turn, the white player performed the 'a' action and the 'black' player action 'd'. How many fluents are true in the second turn? (correct answer: 1, why?)
  7. what is the smallest number of turns required to end the game? (correct answer: 4, why?)
  8. does the game always end? (is it possible to loop the game indefinitely?)

3.2 Assignment 2

Please model the 'crosses and crosses' game. The rules are similar to the tic-tac-toe, again we have nine cells, but both players use the crosses. Player loses when he first creates the line.

3.3 Assignment 3

Please model the “Rock Paper Scissors Lizard Spock” game. You can learn the rules from this video. The image on the right is also a good reference. You can practice it with a colleague or like a modern man: online....

4 Knowledge Interchange Format

Everything we learned so far is true, but the GDL has an alternative syntax used to store and transfer the games: the Knowledge Interchange Format. It's based on a lisp and uses a prefix notation. Moreover :- operator is replaced with , & with and, negation ~ with not. The variables start with the ?. Below you can see an example of translation between Prolog:



and KIF:



4.1 Assignemnt 1

For every example below, tell if the KIF version is a faithful translation of the Prolog one.






4.2 Assignment 2

Download the KIF version of the tic-tac-toe game. What are the differences between it and our model?

4.3 Assignment 3

Read the simple blocks' world model.

  • what are the ending contitions?
  • how do we count the turns?

4.4 Assignment 4

Translate all the models you've created today to the KIF format.

5 Validation

The instructions below require an Eclipse IDE.

Please run the Eclipse IDE and:

  • File → Import Project
  • Import the git repository
    • If there is no such option (Eclipse is too old), one has to download the repository by hand and ignore the further instructions.
  • Clone from the remote URI
  • Select the master branch
  • Select a reasonable catalogue for the project
  • Select a ggp-base project to be imported

5.1 Adding Custom Games

In the project directory go to the games/games catalogue and create there three directories tic-tac-toe, cross-cross and spock. Every directory should contain two files:

  • <name>.kif file with the model
  • METADATA file containing:
{
  "gameName": "<game name>",  
  "rulesheet": "<kif file name>"
}

5.2 Games' Validation

From the Eclipse run the Validator app (one of the targets next to the play button).

Validator will check if the model is a valid GDL model.

Assignment

  1. In the Repository field choose Local Game Repository
  2. In the Game field select Tic-Tac-Toe
  3. Press Validate — There should be no failure
  4. Repeat the same for your custom models
  5. Fix your models to be playable

6 Fun

Run the Kiosk app in the Eclipse (select from the list next to the play button).

Assignment

  1. As an Opponent select SimpleMonteCarloPlayer
  2. Choose a game you are familiar with
  3. Win!

en/dydaktyka/ggp/gdl.1546551800.txt.gz · Last modified: 2019/06/27 16:00 (external edit)
www.chimeric.de Valid CSS Driven by DokuWiki do yourself a favour and use a real browser - get firefox!! Recent changes RSS feed Valid XHTML 1.0