Skip to content

rossjillian/afg

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

afg: AI for Games

Table of Contents

Overview

afg is a template library that aims to make game development in C++ more enjoyable and feasible. It offers powerful code abstractions from the fields of game development, adversarial AI, and model checking to enable game devs to write games from scratch faster with rich player and testing functionality. afg embraces compile-time polymorphism and presents a range of concepts for users to satisfy as opposed to offering abstract base classes for run-time polymorphism. Combined with modern C++ features, these design principles are meant to encourage developers to learn and try to use these elegant C++ constructs in their own implementations!

Game Abstractions

Game devs should simply focus on implementing their own game-specific mechanics and not waste time on boilerplate code. All simple two-player adversarial games share the same mechanics in the end – there is a notion of moves, players, and updating the game state based on the results of the moves those players make. afg acts as the game driver and simply asks the game dev to provide the game mechanics.

Adversarial Artificial Intelligence

Not all games involve human players. afg wants to impower game devs to offer interesting game mechanics by providing the necessary functionality to have AI players to compete against. Of course, we can't expect all game devs to be familiar with adversarial AI, so we instead offer implementations of these algorithms for the game dev to seamlessly plug into their game implemenation! We mainly focused on implementing Minimax and some of its variations.

Model Checking

Finally, game devs need a way to assert the validity of their implementation. Have they covered all edge cases and winning conditions? It would be infeasible for the game dev to have to manually play out the game to reach all of these cases to see if they implementated them correctly. afg solves this issue by providing simple model checking tools for the game dev to use! Through simple predicate-based BFS, game devs will be able to concisely write validity test for their game mechanics.

Design Choices

Intuitive Components

After implementing a few games ourselves, we studied the code and figured out what pieces of the games can be considered as "boilerplate" and therefore abstracted away. We also studied the components that make up a game and used them to implement concepts for Game and Player.

Compile-time Polymorphism

When abstracting away functionality, one may choose to use virtual classes to offer a sort of "skeleton" code for the user to extend. While the OO approach certainly has its use-cases, the penalty of run-time polymorphism is noticable. In afg, we chose to avoid virtual classes in favor of templates, as to validate the polymorphism during compile-time. This yields faster runtime and also means we can take advantage of C++20's Concepts for better type-checking!

Installation

Since afg is a header-only template library, there isn't anything to build! Just pass the compiler flag -I<path> when building your game to include the path to afg/include.

afg requires some C++20 support (mainly for concepts). See here if you are unsure if your system's compiler will support afg. clang++-10 or g++-10 are sufficient.

Tutorial

Game

Let's say we wanted to use afg's game abstractions to implement a simple command-line TicTacToe. We'll first create a TicTacToe struct that complies with the requirements of an afg game. That is we must have the functions:

  • bool isTerminal() - this function should return whether the game has reached a terminal state - for example if someone has won or a tie was reached.
  • bool isWinner() - similar, but only for winners
  • int getTurnCount() - this function should return how many turns into the game we are
  • int getTurnParity() - should simply return which player's turn it is
  • void makeMove(mv) - this function should change whatever internal state you are using to keep track of the game and make the given move
  • void isValid(mv) - this function should check whether a given move is valid
  • void setup() - whatever you want to run at the beginning of the game (potentially instructions, etc) should go here

The implementations of most of these functions are straightforward. We'll focus on isValid(), makeMove() and isWinner(). For the sake of our TicTacToe game, we will create a data structure called a board, which is a wrapper on a vector of vectors of chars which we'll refer to as a grid. In TicTacToe/board.cpp we implemented a bunch of functions We'll need like some functions to print a TicTacToe board on the command line. We also implement the isValid() function:

bool Board::isValid(int move) const
{
   int max = numTiles - 1;
   if (move > max)
      return false;

   int row = move / board.size();
   int col = move % board.size();

   return board[row][col] == ' ';
}

which just checks that a proposed move is within the bounds of the board and is on an empty square.

And we implement the makeMove() function:

void Board::makeMove(int move, int turn, Grid& board){
   int row = move / board.size();
   int col = move % board.size();
   if (isValid(move))
      board[row][col] = turn ? 'x' : 'o';
}

which checks if a move is valid, and if so, changes the symbol in that square to the appropriate one given which player's turn it is.

Finally, we implement an isWinner() function:

bool Board::isWinner() const
{
   return checkRows() || checkColumns() || checkDiagonals();
}

which employs three helper function to check if someone has won on the rows, columns or diagonals.

Once we've written these functions, we simply wrap them in functions in our TicTacToe struct, and we satisfy the constraints of an afg game. Then our TicTacToe game looks very simple:

int main(int argc, char **argv)
{
   HumanPlayer<TicTacToe> p1(0);
   HumanPlayer<TicTacToe> p2(1);

   TicTacToe ttt(3);

   TPGame<TicTacToe, HumanPlayer<TicTacToe>, HumanPlayer<TicTacToe>> game(ttt, p1, p2);

   game.play();

   return 0;
}

Note that we provide two basic players, HumanPlayer and IntelligentPlayer. HumanPlayer can be used out-of-the-box and does not require any additional coding on the part of the game developer. Using HumanPlayer, the game accepts moves from standard input on the command line. All of the looping and setup for the game itself is done by afg (using the functions we wrote). The final game looks like:

Refer to moves using the following chart: 
0 | 1 | 2
------------
3 | 4 | 5
------------
6 | 7 | 8
   |   |  
------------
   |   |  
------------
   |   |  

[ Turn 0 ] Player 0 make a move!
Tile #: 4
   |   |  
------------
   | o |  
------------
   |   |  

[ Turn 1 ] Player 1 make a move!
Tile #: 3
   |   |  
------------
 x | o |  
------------
   |   |  

[ Turn 2 ] Player 0 make a move!
Tile #: 5
   |   |  
------------
 x | o | o
------------
   |   |  

[ Turn 3 ] Player 1 make a move!
Tile #: 7
   |   |  
------------
 x | o | o
------------
   | x |  

[ Turn 4 ] Player 0 make a move!
Tile #:

And so on...

The full code for TicTacToe can be found here. We implemented this game for grids of all sizes, too!

AI

Now that you have your game set up, you probably want to make it a bit more interesting! This is where our AI library comes in. We provide you with a minimax implementation with alpha-beta pruning and a minimax implementation with alpha-beta pruning and iterative deepening. You should consider using our implementation with iterative deepning if the state space of your game is particularly large.

  afg::AI::minimax(state, player, depth);

Here, state is just a const GameType&, which is the GameType you implemented previously.

In Game, you have to enforce the Player concept; this is alreay done for you in the provided HumanPlayer and SmartPlayer. To use our AI functions, you have to additionally enforce the IntelligentPlayer concept, which requires an additional function called the heuristic function. The heuristic function is game-specific and is the evaluation of a state; it is up to you as the game developer to define. All that we require is that the heuristic function should return an integer.

In our TicTacToe implementation, you can see that we have implemented a very simple heuristic function. This heuristic function takes advantage of the zero sum nature of TicTacToe (so each player can maximize its own score).

  template<>
  int SmartPlayer<TicTacToe>::heuristic(const TicTacToe& state) {
      if (state.isWinner()) {
          if (this->getParity() == state.getTurnParity())
              return MAXIMIZER;
          else
              return MINIMIZER;    
      }
      return NEUTRAL;
  }

Model Checking

With TicTacToe pretty much implemented, you may now wish to write code that will verify your implementation. You may also wish to explore some scenarios in the game but don't want to sit down and manually play the game to reach those scenarios. For testing and playing out scenarios, we now arrive to the final component of afg – model checking!

See TicTacToe/model-check.cpp for the full code.

We must first make sure that TicTacToe satisfies the afg::model::Checkable concept. There's a lot of overlap with afg::game::Playable, except we'll also have to implement std::hash<TicTacToe> and operator== for TicTacToe. This is so the game state can be hashed into data structures used by the model checking algorithms.

Once that's out of the way, we're ready to implement some simple model checking for TicTacToe! Let's choose two simple scenarios (predicates) to search for.

First, let's find a game board that looks like this:

  o | o | ?
-------------
  ? | o | ?
-------------
  ? | ? | ?

That is, there must be o's in those three slots, but all the other slots don't matter. We can verify that at least one state like this exists by running

  auto res = afg::model::bfsFind(ttt, threeOs, 10);

Here, ttt is the initial state for TicTacToe, 10 is the depth limit, and threeOs is a lambda that returns true if it sees a state that matches the one above.

  auto threeOs = [](const TicTacToe& st) {
      return (st.b.board[0][0] == 'o'
              && st.b.board[0][1] == 'o'
              && st.b.board[1][1] == 'o');
  }

We now define one more predicate to test if the o-player won:

  auto oWon = [](const TicTacToe& st) {
      return st.b.isWinner() && (st.getTurnParity() == 0); 
  }

With these two predicates, we can define a search to see if a path exists where we first reach a board configuration specified by threeOs and then go on to see the o-player win the game. By specifying a depth limit, we can find the minimum number of turns required to see this scenario happen. After putting the two predicates in a vector, we then call

  bool exists = afg::model::pathExists(ttt, predicates, 7);

This line of code will see if the two predicates can happen within 7 moves.

The goal-based BFS searches are simple, but powerful tools when it comes to model checking complex game states. They make for great testing too!

Benchmarks

See here for more details. At a high-level, we made sure that we were not incurring overhead by mandating that a user hook in with our abstractions instead of just implementing everything themselves. Surely enough, afg follows through with the C++ principle of "Zero Cost Abstraction"!

Manual

For in-depth documentation of afg, see here

Contributors

We made this for our final project in Bjarne Stroustrup's Design Using C++ course at Columbia University.

We drew inspriation from some of our previous coursework at Columbia, including:

  • cs4701 Artificial Intelligence – for our implementations of adversarial AI (Minimax algorithm)
  • cs4113 Distributed Systems – for our use of model checking in asserting correctness of game mechanics

About

afg: AI for Games

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •