GAMES OR ADVERSARIAL SEARCH

This notebook serves as supporting material for topics covered in Chapter 5 - Adversarial Search in the book Artificial Intelligence: A Modern Approach. This notebook uses implementations from games.py module. Let's import required classes, methods, global variables etc., from games module.

Game Representation
Game Examples
- Tic-Tac-Toe
- Figure 5.2 Game
Min-Max
Alpha-Beta
Players
Let's Play Some Games!



In [1]:

    
from games import *
from notebook import psource, pseudocode

GAME REPRESENTATION

To represent games we make use of the Game class, which we can subclass and override its functions to represent our own games. A helper tool is the namedtuple GameState, which in some cases can come in handy, especially when our game needs us to remember a board (like chess).

`GameState` namedtuple

GameState is a namedtuple which represents the current state of a game. It is used to help represent games whose states can't be easily represented normally, or for games that require memory of a board, like Tic-Tac-Toe.

Gamestate is defined as follows:

GameState = namedtuple('GameState', 'to_move, utility, board, moves')

to_move: It represents whose turn it is to move next.
utility: It stores the utility of the game state. Storing this utility is a good idea, because, when you do a Minimax Search or an Alphabeta Search, you generate many recursive calls, which travel all the way down to the terminal states. When these recursive calls go back up to the original callee, we have calculated utilities for many game states. We store these utilities in their respective GameStates to avoid calculating them all over again.
board: A dict that stores the board of the game.
moves: It stores the list of legal moves possible from the current position.

`Game` class

Let's have a look at the class Game in our module. We see that it has functions, namely actions, result, utility, terminal_test, to_move and display.

We see that these functions have not actually been implemented. This class is just a template class; we are supposed to create the class for our game, by inheriting this Game class and implementing all the methods mentioned in Game.



In [2]:

    
%psource Game

Now let's get into details of all the methods in our Game class. You have to implement these methods when you create new classes that would represent your game.

actions(self, state): Given a game state, this method generates all the legal actions possible from this state, as a list or a generator. Returning a generator rather than a list has the advantage that it saves space and you can still operate on it as a list.

result(self, state, move): Given a game state and a move, this method returns the game state that you get by making that move on this game state.

utility(self, state, player): Given a terminal game state and a player, this method returns the utility for that player in the given terminal game state. While implementing this method assume that the game state is a terminal game state. The logic in this module is such that this method will be called only on terminal game states.

terminal_test(self, state): Given a game state, this method should return True if this game state is a terminal state, and False otherwise.

to_move(self, state): Given a game state, this method returns the player who is to play next. This information is typically stored in the game state, so all this method does is extract this information and return it.

display(self, state): This method prints/displays the current state of the game.

GAME EXAMPLES

Below we give some examples for games you can create and experiment on.

Tic-Tac-Toe

Take a look at the class TicTacToe. All the methods mentioned in the class Game have been implemented here.



In [3]:

    
%psource TicTacToe

The class TicTacToe has been inherited from the class Game. As mentioned earlier, you really want to do this. Catching bugs and errors becomes a whole lot easier.

Additional methods in TicTacToe:

__init__(self, h=3, v=3, k=3) : When you create a class inherited from the Game class (class TicTacToe in our case), you'll have to create an object of this inherited class to initialize the game. This initialization might require some additional information which would be passed to __init__ as variables. For the case of our TicTacToe game, this additional information would be the number of rows h, number of columns v and how many consecutive X's or O's are needed in a row, column or diagonal for a win k. Also, the initial game state has to be defined here in __init__.

compute_utility(self, board, move, player) : A method to calculate the utility of TicTacToe game. If 'X' wins with this move, this method returns 1; if 'O' wins return -1; else return 0.

k_in_row(self, board, move, player, delta_x_y) : This method returns True if there is a line formed on TicTacToe board with the latest move else False.

TicTacToe GameState

Now, before we start implementing our TicTacToe game, we need to decide how we will be representing our game state. Typically, a game state will give you all the current information about the game at any point in time. When you are given a game state, you should be able to tell whose turn it is next, how the game will look like on a real-life board (if it has one) etc. A game state need not include the history of the game. If you can play the game further given a game state, you game state representation is acceptable. While we might like to include all kinds of information in our game state, we wouldn't want to put too much information into it. Modifying this game state to generate a new one would be a real pain then.

Now, as for our TicTacToe game state, would storing only the positions of all the X's and O's be sufficient to represent all the game information at that point in time? Well, does it tell us whose turn it is next? Looking at the 'X's and O's on the board and counting them should tell us that. But that would mean extra computing. To avoid this, we will also store whose move it is next in the game state.

Think about what we've done here. We have reduced extra computation by storing additional information in a game state. Now, this information might not be absolutely essential to tell us about the state of the game, but it does save us additional computation time. We'll do more of this later on.

To store game states will will use the GameState namedtuple.

to_move: A string of a single character, either 'X' or 'O'.
utility: 1 for win, -1 for loss, 0 otherwise.
board: All the positions of X's and O's on the board.
moves: All the possible moves from the current state. Note here, that storing the moves as a list, as it is done here, increases the space complexity of Minimax Search from O(m) to O(bm). Refer to Sec. 5.2.1 of the book.

Representing a move in TicTacToe game

Now that we have decided how our game state will be represented, it's time to decide how our move will be represented. Becomes easy to use this move to modify a current game state to generate a new one.

For our TicTacToe game, we'll just represent a move by a tuple, where the first and the second elements of the tuple will represent the row and column, respectively, where the next move is to be made. Whether to make an 'X' or an 'O' will be decided by the to_move in the GameState namedtuple.

Fig52 Game

For a more trivial example we will represent the game in Figure 5.2 of the book.

The states are represented wih capital letters inside the triangles (eg. "A") while moves are the labels on the edges between states (eg. "a1"). Terminal nodes carry utility values. Note that the terminal nodes are named in this example 'B1', 'B2' and 'B2' for the nodes below 'B', and so forth.

We will model the moves, utilities and initial state like this:



In [4]:

    
moves = dict(A=dict(a1='B', a2='C', a3='D'),
                 B=dict(b1='B1', b2='B2', b3='B3'),
                 C=dict(c1='C1', c2='C2', c3='C3'),
                 D=dict(d1='D1', d2='D2', d3='D3'))
utils = dict(B1=3, B2=12, B3=8, C1=2, C2=4, C3=6, D1=14, D2=5, D3=2)
initial = 'A'

In moves, we have a nested dictionary system. The outer's dictionary has keys as the states and values the possible moves from that state (as a dictionary). The inner dictionary of moves has keys the move names and values the next state after the move is complete.

Below is an example that showcases moves. We want the next state after move 'a1' from 'A', which is 'B'. A quick glance at the above image confirms that this is indeed the case.



In [5]:

    
print(moves['A']['a1'])

We will now take a look at the functions we need to implement. First we need to create an object of the Fig52Game class.



In [6]:

    
fig52 = Fig52Game()

actions: Returns the list of moves one can make from a given state.



In [ ]:

    
psource(Fig52Game.actions)



In [8]:

    
print(fig52.actions('B'))









    



['b1', 'b2', 'b3']

result: Returns the next state after we make a specific move.



In [ ]:

    
psource(Fig52Game.result)



In [10]:

    
print(fig52.result('A', 'a1'))

utility: Returns the value of the terminal state for a player ('MAX' and 'MIN'). Note that for 'MIN' the value returned is the negative of the utility.



In [ ]:

    
psource(Fig52Game.utility)



In [12]:

    
print(fig52.utility('B1', 'MAX'))
print(fig52.utility('B1', 'MIN'))

terminal_test: Returns True if the given state is a terminal state, False otherwise.



In [ ]:

    
psource(Fig52Game.terminal_test)



In [14]:

    
print(fig52.terminal_test('C3'))









    



True

to_move: Return the player who will move in this state.



In [ ]:

    
psource(Fig52Game.to_move)



In [16]:

    
print(fig52.to_move('A'))

MAX

As a whole the class Fig52 that inherits from the class Game and overrides its functions:



In [ ]:

    
psource(Fig52Game)

MIN-MAX

Overview

This algorithm (often called Minimax) computes the next move for a player (MIN or MAX) at their current state. It recursively computes the minimax value of successor states, until it reaches terminals (the leaves of the tree). Using the utility value of the terminal states, it computes the values of parent states until it reaches the initial node (the root of the tree).

It is worth noting that the algorithm works in a depth-first manner. The pseudocode can be found below:



In [2]:

    
pseudocode("Minimax-Decision")









    Out[2]:




AIMA3e
function MINIMAX-DECISION(state) returns an action

 return arg max_{_a_ &Element ACTIONS(_s_)} MIN-VALUE(RESULT(state, a))

function MAX-VALUE(state) returns a utility value

 if TERMINAL-TEST(state) then return UTILITY(state)

 v ← −∞

 for each a in ACTIONS(state) do

   v ← MAX(v, MIN-VALUE(RESULT(state, a)))

 return v

function MIN-VALUE(state) returns a utility value

 if TERMINAL-TEST(state) then return UTILITY(state)

 v ← ∞

 for each a in ACTIONS(state) do

   v ← MIN(v, MAX-VALUE(RESULT(state, a)))

 return v

Figure ?? An algorithm for calculating minimax decisions. It returns the action corresponding to the best possible move, that is, the move that leads to the outcome with the best utility, under the assumption that the opponent plays to minimize utility. The functions MAX-VALUE and MIN-VALUE go through the whole game tree, all the way to the leaves, to determine the backed-up value of a state. The notation argmax _{_a_ &Element _S_} f(a) computes the element a of set S that has maximum value of f(a).

Implementation

In the implementation we are using two functions, max_value and min_value to calculate the best move for MAX and MIN respectively. These functions interact in an alternating recursion; one calls the other until a terminal state is reached. When the recursion halts, we are left with scores for each move. We return the max. Despite returning the max, it will work for MIN too since for MIN the values are their negative (hence the order of values is reversed, so the higher the better for MIN too).



In [ ]:

    
psource(minimax_decision)

Example

We will now play the Fig52 game using this algorithm. Take a look at the Fig52Game from above to follow along.

It is the turn of MAX to move, and he is at state A. He can move to B, C or D, using moves a1, a2 and a3 respectively. MAX's goal is to maximize the end value. So, to make a decision, MAX needs to know the values at the aforementioned nodes and pick the greatest one. After MAX, it is MIN's turn to play. So MAX wants to know what will the values of B, C and D be after MIN plays.

The problem then becomes what move will MIN make at B, C and D. The successor states of all these nodes are terminal states, so MIN will pick the smallest value for each node. So, for B he will pick 3 (from move b1), for C he will pick 2 (from move c1) and for D he will again pick 2 (from move d3).

Let's see this in code:



In [19]:

    
print(minimax_decision('B', fig52))
print(minimax_decision('C', fig52))
print(minimax_decision('D', fig52))

Now MAX knows that the values for B, C and D are 3, 2 and 2 (produced by the above moves of MIN). The greatest is 3, which he will get with move a1. This is then the move MAX will make. Let's see the algorithm in full action:



In [20]:

    
print(minimax_decision('A', fig52))

a1

Visualization

Below we have a simple game visualization using the algorithm. After you run the command, click on the cell to move the game along. You can input your own values via a list of 27 integers.



In [2]:

    
from notebook import Canvas_minimax
from random import randint



In [ ]:

    
minimax_viz = Canvas_minimax('minimax_viz', [randint(1, 50) for i in range(27)])

ALPHA-BETA

Overview

While Minimax is great for computing a move, it can get tricky when the number of game states gets bigger. The algorithm needs to search all the leaves of the tree, which increase exponentially to its depth.

For Tic-Tac-Toe, where the depth of the tree is 9 (after the 9th move, the game ends), we can have at most 9! terminal states (at most because not all terminal nodes are at the last level of the tree; some are higher up because the game ended before the 9th move). This isn't so bad, but for more complex problems like chess, we have over $10^{40}$ terminal nodes. Unfortunately we have not found a way to cut the exponent away, but we nevertheless have found ways to alleviate the workload.

Here we examine pruning the game tree, which means removing parts of it that we do not need to examine. The particular type of pruning is called alpha-beta, and the search in whole is called alpha-beta search.

To showcase what parts of the tree we don't need to search, we will take a look at the example Fig52Game.

In the example game, we need to find the best move for player MAX at state A, which is the maximum value of MIN's possible moves at successor states.

MAX(A) = MAX( MIN(B), MIN(C), MIN(D) )

MIN(B) is the minimum of 3, 12, 8 which is 3. So the above formula becomes:

MAX(A) = MAX( 3, MIN(C), MIN(D) )

Next move we will check is c1, which leads to a terminal state with utility of 2. Before we continue searching under state C, let's pop back into our formula with the new value:

MAX(A) = MAX( 3, MIN(2, c2, .... cN), MIN(D) )

We do not know how many moves state C allows, but we know that the first one results in a value of 2. Do we need to keep searching under C? The answer is no. The value MIN will pick on C will at most be 2. Since MAX already has the option to pick something greater than that, 3 from B, he does not need to keep searching under C.

In alpha-beta we make use of two additional parameters for each state/node, a and b, that describe bounds on the possible moves. The parameter a denotes the best choice (highest value) for MAX along that path, while b denotes the best choice (lowest value) for MIN. As we go along we update a and b and prune a node branch when the value of the node is worse than the value of a and b for MAX and MIN respectively.

In the above example, after the search under state B, MAX had an a value of 3. So, when searching node C we found a value less than that, 2, we stopped searching under C.

You can read the pseudocode below:



In [3]:

    
pseudocode("Alpha-Beta-Search")









    Out[3]:




AIMA3e
function ALPHA-BETA-SEARCH(state) returns an action

 v ← MAX-VALUE(state, −∞, &plus∞)

 return the action in ACTIONS(state) with value v

function MAX-VALUE(state, α, β) returns a utility value

 if TERMINAL-TEST(state) then return UTILITY(state)

 v ← −∞

 for each a in ACTIONS(state) do

   v ← MAX(v, MIN-VALUE(RESULT(state, a), α, β))

   if v ≥ β then return v

   α ← MAX(α, v)

 return v

function MIN-VALUE(state, α, β) returns a utility value

 if TERMINAL-TEST(state) then return UTILITY(state)

 v ← &plus∞

 for each a in ACTIONS(state) do

   v ← MIN(v, MAX-VALUE(RESULT(state, a), α, β))

   if v ≤ α then return v

   β ← MIN(β, v)

 return v

Figure ?? The alpha-beta search algorithm. Notice that these routines are the same as the MINIMAX functions in Figure ??, except for the two lines in each of MIN-VALUE and MAX-VALUE that maintain α and β (and the bookkeeping to pass these parameters along).

Implementation

Like minimax, we again make use of functions max_value and min_value, but this time we utilise the a and b values, updating them and stopping the recursive call if we end up on nodes with values worse than a and b (for MAX and MIN). The algorithm finds the maximum value and returns the move that results in it.

The implementation:



In [21]:

    
%psource alphabeta_search

Example

We will play the Fig52 Game with the alpha-beta search algorithm. It is the turn of MAX to play at state A.



In [22]:

    
print(alphabeta_search('A', fig52))

a1

The optimal move for MAX is a1, for the reasons given above. MIN will pick move b1 for B resulting in a value of 3, updating the a value of MAX to 3. Then, when we find under C a node of value 2, we will stop searching under that sub-tree since it is less than a. From D we have a value of 2. So, the best move for MAX is the one resulting in a value of 3, which is a1.

Below we see the best moves for MIN starting from B, C and D respectively. Note that the algorithm in these cases works the same way as minimax, since all the nodes below the aforementioned states are terminal.



In [23]:

    
print(alphabeta_search('B', fig52))
print(alphabeta_search('C', fig52))
print(alphabeta_search('D', fig52))

Visualization

Below you will find the visualization of the alpha-beta algorithm for a simple game. Click on the cell after you run the command to move the game along. You can input your own values via a list of 27 integers.



In [2]:

    
from notebook import Canvas_alphabeta
from random import randint



In [ ]:

    
alphabeta_viz = Canvas_alphabeta('alphabeta_viz', [randint(1, 50) for i in range(27)])

PLAYERS

So, we have finished the implementation of the TicTacToe and Fig52Game classes. What these classes do is defining the rules of the games. We need more to create an AI that can actually play games. This is where random_player and alphabeta_player come in.

query_player

The query_player function allows you, a human opponent, to play the game. This function requires a display method to be implemented in your game class, so that successive game states can be displayed on the terminal, making it easier for you to visualize the game and play accordingly.

random_player

The random_player is a function that plays random moves in the game. That's it. There isn't much more to this guy.

alphabeta_player

The alphabeta_player, on the other hand, calls the alphabeta_search function, which returns the best move in the current game state. Thus, the alphabeta_player always plays the best move given a game state, assuming that the game tree is small enough to search entirely.

play_game

The play_game function will be the one that will actually be used to play the game. You pass as arguments to it an instance of the game you want to play and the players you want in this game. Use it to play AI vs AI, AI vs human, or even human vs human matches!

LET'S PLAY SOME GAMES!

Game52

Let's start by experimenting with the Fig52Game first. For that we'll create an instance of the subclass Fig52Game inherited from the class Game:



In [27]:

    
game52 = Fig52Game()

First we try out our random_player(game, state). Given a game state it will give us a random move every time:



In [28]:

    
print(random_player(game52, 'A'))
print(random_player(game52, 'A'))

The alphabeta_player(game, state) will always give us the best move possible, for the relevant player (MAX or MIN):



In [29]:

    
print( alphabeta_player(game52, 'A') )
print( alphabeta_player(game52, 'B') )
print( alphabeta_player(game52, 'C') )

What the alphabeta_player does is, it simply calls the method alphabeta_full_search. They both are essentially the same. In the module, both alphabeta_full_search and minimax_decision have been implemented. They both do the same job and return the same thing, which is, the best move in the current state. It's just that alphabeta_full_search is more efficient with regards to time because it prunes the search tree and hence, explores lesser number of states.



In [30]:

    
minimax_decision('A', game52)









    Out[30]:





'a1'



In [31]:

    
alphabeta_search('A', game52)









    Out[31]:





'a1'

Demonstrating the play_game function on the game52:



In [32]:

    
game52.play_game(alphabeta_player, alphabeta_player)



In [33]:

    
game52.play_game(alphabeta_player, random_player)



In [34]:

    
game52.play_game(query_player, alphabeta_player)









    



current state:
A
available moves: ['a1', 'a2', 'a3']

Your move? a1
B1






    Out[34]:





3



In [35]:

    
game52.play_game(alphabeta_player, query_player)









    



current state:
B
available moves: ['b1', 'b2', 'b3']

Your move? b1
B1






    Out[35]:





3

Note that if you are the first player then alphabeta_player plays as MIN, and if you are the second player then alphabeta_player plays as MAX. This happens because that's the way the game is defined in the class Fig52Game. Having a look at the code of this class should make it clear.

TicTacToe

Now let's play TicTacToe. First we initialize the game by creating an instance of the subclass TicTacToe inherited from the class Game:



In [36]:

    
ttt = TicTacToe()

We can print a state using the display method:



In [37]:

    
ttt.display(ttt.initial)









    



. . . 
. . . 
. . .

Hmm, so that's the initial state of the game; no X's and no O's.

Let us create a new game state by ourselves to experiment:



In [38]:

    
my_state = GameState(
    to_move = 'X',
    utility = '0',
    board = {(1,1): 'X', (1,2): 'O', (1,3): 'X',
             (2,1): 'O',             (2,3): 'O',
             (3,1): 'X',
            },
    moves = [(2,2), (3,2), (3,3)]
    )

So, how does this game state look like?



In [39]:

    
ttt.display(my_state)









    



X O X 
O . O 
X . .

The random_player will behave how he is supposed to i.e. pseudo-randomly:



In [40]:

    
random_player(ttt, my_state)









    Out[40]:





(2, 2)



In [41]:

    
random_player(ttt, my_state)









    Out[41]:





(2, 2)

But the alphabeta_player will always give the best move, as expected:



In [42]:

    
alphabeta_player(ttt, my_state)









    Out[42]:





(2, 2)

Now let's make two players play against each other. We use the play_game function for this. The play_game function makes players play the match against each other and returns the utility for the first player, of the terminal state reached when the game ends. Hence, for our TicTacToe game, if we get the output +1, the first player wins, -1 if the second player wins, and 0 if the match ends in a draw.



In [43]:

    
ttt.play_game(random_player, alphabeta_player)









    



O O . 
X O X 
X X O 






    Out[43]:





-1

The output is (usually) -1, because random_player loses to alphabeta_player. Sometimes, however, random_player manages to draw with alphabeta_player.

Since an alphabeta_player plays perfectly, a match between two alphabeta_players should always end in a draw. Let's see if this happens:



In [44]:

    
for _ in range(10):
    print(ttt.play_game(alphabeta_player, alphabeta_player))









    



X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0
X X O 
O O X 
X O X 
0

A random_player should never win against an alphabeta_player. Let's test that.



In [45]:

    
for _ in range(10):
    print(ttt.play_game(random_player, alphabeta_player))









    



X O O 
X O . 
O X X 
-1
O X . 
O X X 
O . . 
-1
X X O 
O O X 
O X . 
-1
O O O 
. X X 
X . . 
-1
O O O 
. . X 
X . X 
-1
O X O 
X O X 
X . O 
-1
O X X 
O X X 
O O . 
-1
O O X 
X O X 
X O . 
-1
O O X 
X O . 
X O X 
-1
O O X 
X X O 
O X X 
0

Canvas_TicTacToe(Canvas)

This subclass is used to play TicTacToe game interactively in Jupyter notebooks. TicTacToe class is called while initializing this subclass.

Let's have a match between random_player and alphabeta_player. Click on the board to call players to make a move.



In [46]:

    
from notebook import Canvas_TicTacToe



In [47]:

    
bot_play = Canvas_TicTacToe('bot_play', 'random', 'alphabeta')

Now, let's play a game ourselves against a random_player:



In [48]:

    
rand_play = Canvas_TicTacToe('rand_play', 'human', 'random')

Yay! We (usually) win. But we cannot win against an alphabeta_player, however hard we try.



In [49]:

    
ab_play = Canvas_TicTacToe('ab_play', 'human', 'alphabeta')

GAMES OR ADVERSARIAL SEARCH

CONTENTS

GAME REPRESENTATION

GameState namedtuple

Game class

GAME EXAMPLES

Tic-Tac-Toe

TicTacToe GameState

Representing a move in TicTacToe game

Fig52 Game

MIN-MAX

Overview

AIMA3e

Implementation

Example

Visualization

ALPHA-BETA

Overview

AIMA3e

Implementation

Example

Visualization

PLAYERS

query_player

random_player

alphabeta_player

play_game

LET'S PLAY SOME GAMES!

Game52

TicTacToe

Canvas_TicTacToe(Canvas)

`GameState` namedtuple

`Game` class