Home

“It was a watershed event, but it doesn’t have to do with computers becoming intelligent,” said Douglas Hofstadter, a professor of computer science at Indiana University and author of several books about human intelligence, including Gödel, Escher, Bach, which won a Pulitzer Prize in 1980, with its witty argument about the connecting threads of intellect in various fields of expression. “They’re just overtaking humans in certain intellectual activities that we thought required intelligence. My God, I used to think chess required thought. Now, I realize it doesn’t. It doesn’t mean Kasparov isn’t a deep thinker, just that you can bypass deep thinking in playing chess, the way you can fly without flapping your wings.” — Bruce Weber, “Mean Chess-Playing Computer Tears at Meaning of Thought,” The New York Times, Feb 19, 1996

Chess, tic-tac-toe, connect-four, checkers, go, etc. etc. are zero-sum games, meaning in the end, the gains and losses among the players add up to zero. In the examples listed, one player wins, another loses, so the sum of “utility” for each player at the end is zero.

Poker is another zero-sum game, although there is one winner and many losers. However, the amount the winner wins is equal to the sum of the amounts the losers lose.

Study of zero-sum games began the study of game theory, which is a mathematical subject that covers any situation involving several agents attempting to make decisions that maximize their own utility.

We will not be studying game theory, as a whole, but we will learn how to “play” zero-sum games.

## Expansion of the search problem

The components of a search problem included the following:

Initial state
some description of the agent’s starting situation
Possible actions
the set of actions (such as chess moves) available to the agent, also called “applicable” actions; the possible actions depend on the state
Transition model
some way of figuring out what an action does; in other words, a `resultOf(state, action)` function which returns a state; the transition model defines a state space, which takes the form of a directed graph (vertices are states, edges are actions)

Note that we have left out “Goal criteria” and “Path cost.” In order to support adversarial search, we add the following:

Players
a list of players (we’ll look at 2-player games) and a `switch_player(player)` function that switches to the player following `player`
Terminal tests
a function or functions that test for final states (winning/losing/tied states)
Utility function
a function `utility(state, player)` that returns a real number (or integer) representing the value to `player` of `state`; presumably, winning states have the highest utility, losing states the lowest utility, tied states have zero utility, and states that are not terminal have a utility calculated by looking at the utilities of following states

## I move, you move, etc.

The simplest zero-sum games are those that have two players who take turns. Chess, tic-tac-toe, etc. fit this definition. Suppose a human is playing a computer. The computer needs to figure out which move is best (the computer wants to maximize the utility of its move). We learned how to search (BFS, DFS, A*, and so on), but even though the computer will be engaging in a search, those search algorithms will not help us here.

BFS, DFS, A* and so on attempt to maximize (or minimize) every choice equally (for example, we want to find the shortest route). But in a game search, instead we want to maximize the payoff of our choices but minimize the payoff of the other player’s choices.

The way to do this is to treat the opponent’s positive utility as our negative utility. So if the opponent makes a winning move, that move is worth 1.0 to the opponent but worth -1.0 to us (a tie game has utility 0.0 for both players). We want to find the move that maximizes our utility (best case, 1.0) and minimizes the opponent’s utility (best case, -1.0).

## Minimax algorithm

The minimax algorithm does exactly that: it finds the move that minimizes the maximum utility the opponent can obtain. You can call it “maximin” if you’d like: maximize the minimum utility we (not the opponent) can obtain. They are the same algorithm, just the order of the steps (maximize or minimize) is different.

First, we consider all of our possible moves. If a move is a final move (winning/losing/tie), then we calculate its utility directly (1.0 for a win, -1.0 for a loss, 0.0 for a tie). If it’s not a final move, we have to look further to calculate its utility. We choose one move from which to “look further,” and consider all the possible moves the opponent might make. If the opponent wins, the utility of the opponent’s move is -1.0. Of all the possible moves the opponent might make, we find the worst (minimum) utility. That utility is the value of our move.

Of all our possible moves, we choose the one that has the maximum utility; it is the maximum of the minimum utilities generated by the opponent’s moves. We are assuming the opponent is also choosing the best move, so when we choose our move, we want to find the move that performs best (max) assuming the opponent also performs best (min; i.e., max for the opponent, min for us).

Here is an example search tree for tic-tac-toe. We basically have a depth-first search but also propogate information about utilities (max or min) upwards.

### Python code for minimax

```def minimax(state):
max_trans = None
max_u = None
transitions = possible_transitions(state, 'x')
# Find the transition (move) that provides the maximum
# utility, assuming the opponent also makes a best move
for trans, nextstate in transitions.iteritems():
# after making our move, find the best move the
# opponent can make (best for opponent = worse for us;
# if we consider the opponent winning as negative
# utility, we want to find the minimum utility move
# of the opponent's possible moves)
u = min_utility(nextstate, 'o')
if max_u is None or u > max_u:
max_trans = trans
max_u = u
return max_trans

def min_utility(state, player):
# if the current state is a win/loss/tie, stop searching
if is_winning(state, player) or \
is_winning(state, switch_player(player)) or \
is_tie(state):
return utility(state, 'x')
else:
transitions = possible_transitions(state, player)
min_u = None
for nextstate in transitions.values():
# after making a move (current player is in the
# "player" variable), find the minimum next
# move and return its utility
u = max_utility(nextstate, switch_player(player))
if min_u is None or u < min_u:
min_u = u
return min_u

def max_utility(state, player):
# if the current state is a win/loss/tie, stop searching
if is_winning(state, player) or \
is_winning(state, switch_player(player)) or \
is_tie(state):
return utility(state, 'x')
else:
transitions = possible_transitions(state, player)
max_u = None
for nextstate in transitions.values():
# after making a move (current player is in the
# "player" variable), find the maximum next
# move and return its utility
u = min_utility(nextstate, switch_player(player))
if max_u is None or u > max_u:
max_u = u
return max_u
```

### Python code for tic-tac-toe components

Here is some code for tic-tac-toe specific components used by the minimax algorithm. Not all the code needed to play tic-tac-toe is shown here, however.

```def switch_player(player):
if player == 'x': return 'o'
else: return 'x'

def is_winning(state, player):
winning = False
for i in [0,1,2]:
if state[i] == state[i] == state[i] == player:
winning = True
for j in [0,1,2]:
if state[j] == state[j] == state[j] == player:
winning = True
if state == state == state == player:
winning = True
if state == state == state == player:
winning = True
return winning

def is_tie(state):
blanks = 0
for i in [0,1,2]:
for j in [0,1,2]:
if state[i][j] == ' ': blanks += 1
return(blanks == 0 and \
not is_winning(state, 'x') and \
not is_winning(state, 'o'))

def utility(state, player):
if is_winning(state, player): return 1
if is_winning(state, switch_player(player)): return -1
return 0
```

## Alpha-beta pruning

The minimax search is very costly. Far too many states are searched. We can use a very simple, and somewhat obvious, optimization to search only a fraction as many states while coming up with the exact same answer as minimax.

The key insight is that when we are performing the “max” step (or equivalently, the “min” step), if we find that a certain move gives a maximum utility X (by searching its entire subtree), we can avoid any searches of subtrees whose minimum value is less than X.

Look here, from `min_utility` (similar code is found in `max_utility` and `minimax` functions as well):

```min_u = None
for nextstate in transitions.values():
# after making a move (current player is in the
# "player" variable), find the minimum next
# move and return its utility
u = max_utility(nextstate, switch_player(player))
if min_u is None or u < min_u:
min_u = u
return min_u
```

The variable `min_u` here will only get smaller. Recall that the utility that `min_utility` returns will be used by the `max_utility` function. So different calls to `min_utility` will be compared to pick out the maximum `min_utility` value. Suppose the `min_utilty` function was told (via a parameter) that some other call to `min_utility` produced a minimum utility of α, and that so far, α is the greatest minimum utility found so far (the maximum of the minimums).

Then if `min_u` ever goes smaller than α, there is no reason to continue searching the other transitions (game moves), because `min_u` can only go smaller (being a minimum), and any value smaller than α (the previously-found maximum `min_u` produced by earlier function calls) will simply be ignored when maximizing.

The same can be said about `max_u` in the `max_utility` function; we use the β parameter to keep track of the smallest `max_u` so far seen.

When maximizing (in `max_utility`), we will only change α because we want to know the maximum of the minimums (α is the greatest lower bound); when minimizing (in `min_utility`), we will only change β.

α
the utility of the best (highest-value) choice we have found so far at any choice point along the path in the “max” mode of minimax
β
the utility of the best (lowest-value) choice for the “min” mode of minimax

As we search down the tree, we provide the recursive search procedure our currently-known best α and β. If the recursive search sees that some subtree cannot give any utility better than α (or worse than β) then we don’t bother searching further from there.

It’s very important to realize that the alpha-beta pruning procedure does not produce different answers than minimax. Alpha-beta pruning simply does not bother checking subtrees when it is certain those subtrees will not change the answer produced by minimax.

Check out the final search graph of another tic-tac-toe example, this time with alpha-beta pruning. For comparison, this is the search graph without alpha-beta pruning.

Alpha-beta pruning performs much better than vanilla minimax:

Number of x/o tokens on the boardAverage minimax states checkedAverage α-β states checked
0549946.029019.0
27293.01308.0
4176.281.8
611.09.2
81.31.3

### Python code for alpha-beta pruning

The minimax code above needs few modifications to support alpha-beta pruning:

```def minimax_ab(state):
max_trans = None
max_u = None
alpha = None
beta = None
transitions = possible_transitions(state, 'x')
# Find the transition (move) that provides the maximum
# utility, assuming the opponent also makes a best move
for trans, nextstate in transitions.iteritems():
# after making our move, find the best move the
# opponent can make (best for opponent = worse for us;
# if we consider the opponent winning as negative
# utility, we want to find the minimum utility move
# of the opponent's possible moves)
u = min_utility_ab(nextstate, 'o', alpha, beta)
if max_u is None or u > max_u:
max_trans = trans
max_u = u
# if the utility we just found is greater than beta, and these
# utilities can only get bigger (because we're in a max
# stage), then there is no reason to check further
if beta is not None and u >= beta:
return trans
# if the utility we just found (from the min stage) is greater
# than alpha, we found a new greatest min; this will restrict
# future searches not to look further if they are minimizing
# and they find a utility less than alpha
if alpha is None or u > alpha:
alpha = u
return max_trans

def min_utility_ab(state, player, alpha, beta):
# if the current state is a win/loss/tie, stop searching
if is_winning(state, player) or \
is_winning(state, switch_player(player)) or \
is_tie(state):
return utility(state, 'x')
else:
transitions = possible_transitions(state, player)
min_u = None
for nextstate in transitions.values():
# after making a move (current player is in the
# "player" variable), find the minimum next
# move and return its utility
u = max_utility_ab(nextstate, switch_player(player), alpha, beta)
if min_u is None or u < min_u:
min_u = u
# if the utility we just found is smaller than alpha, and
# these utilities can only get smaller (because we're in a
# min stage), then there is no reason to check further
if alpha is not None and u <= alpha:
return u
# if the utility we just found (from the max stage) is
# smaller than beta, we found a new smallest max; this
# will restrict future searches not to look further if
# they are maximizing and they find a utility greater than
# beta
if beta is None or u < beta:
beta = u
return min_u

def max_utility_ab(state, player, alpha, beta):
# if the current state is a win/loss/tie, stop searching
if is_winning(state, player) or \
is_winning(state, switch_player(player)) or \
is_tie(state):
return utility(state, 'x')
else:
transitions = possible_transitions(state, player)
max_u = None
for nextstate in transitions.values():
# after making a move (current player is in the
# "player" variable), find the maximum next
# move and return its utility
u = min_utility_ab(nextstate, switch_player(player), alpha, beta)
if max_u is None or u > max_u:
max_u = u
# if the utility we just found is greater than beta, and
# these utilities can only get bigger (because we're in a
# max stage), then there is no reason to check further
if beta is not None and u >= beta:
return u
# if the utility we just found (from the min stage) is
# greater than alpha, we found a new greatest min; this
# will restrict future searches not to look further if
# they are minimizing and they find a utility less than
# alpha
if alpha is None or u > alpha:
alpha = u
return max_u
```

## A poor example?

Tic-tac-toe is such a simple game that minimax and alpha-beta pruning are not really necessary. Apparently, here is optimal play (from xkcd): ## I can’t help myself

This is also instructive (from xkcd): 