Negamax

Negamax search is a variant form of minimax search that relies on the zero-sum property of a two-player game.

This algorithm relies on the fact that max(a, b) = −min(−a, −b) to simplify the implementation of the minimax algorithm. More precisely, the value of a position to player A in such a game is the negation of the value to player B. Thus, the player on move looks for a move that maximizes the negation of the value resulting from the move: this successor position must by definition have been valued by the opponent. The reasoning of the previous sentence works regardless of whether A or B is on move. This means that a single procedure can be used to value both positions. This is a coding simplification over minimax, which requires that A selects the move with the maximum-valued successor while B selects the move with the minimum-valued successor.

It should not be confused with negascout, an algorithm to compute the minimax or negamax value quickly by clever use of alpha-beta pruning discovered in the 1980s. Note that alpha-beta pruning is itself a way to compute the minimax or negamax value of a position quickly by avoiding the search of certain uninteresting positions.

Most adversarial search engines are coded using some form of negamax search.

Negamax base algorithm

An animated pedagogical example showing the plain negamax algorithm (that is, without alpha-beta pruning). The person performing the game tree search is considered to be the one that has to move first from the current state of the game (player in this case)

NegaMax operates on the same game trees as those used with the minimax search algorithm. Each node and root node in the tree are game states (such as game board configuration) of a two player game. Transitions to child nodes represent moves available to a player who's about to play from a given node.

The negamax search objective is to find the node score value for the player who is playing at the root node. The pseudocode below shows the negamax base algorithm,[1] with a configurable limit for the maximum search depth:

01 function negamax(node, depth, color)
02     if depth = 0 or node is a terminal node
03         return color * the heuristic value of node

04     bestValue := −∞
05     foreach child of node
06         v := −negamax(child, depth − 1, −color)
07         bestValue := max( bestValue, v )
08     return bestValue
Initial call for Player A's root node
rootNegamaxValue := negamax( rootNode, depth, 1)
rootMinimaxValue := rootNegamaxValue
Initial call for Player B's root node
rootNegamaxValue := negamax( rootNode, depth, −1)
rootMinimaxValue := −rootNegamaxValue

The root node inherits its score from one of its immediate child nodes. The child node that ultimately sets the root node's best score also represents the best move to play. Although the negamax function shown only returns the node's best score as bestValue, practical negamax implementations may also retain and return both best move and best score for the root node. Assuming basic negamax, only the node's best score is essential with non-root nodes. And the node's best move isn't necessary to retain nor return for those nodes.

What can be confusing is how the heuristic value of the current node is calculated. In this implementation, this value is always calculated from the point of view of player A, whose color value is one. In other words, higher heuristic values always represent situations more favorable for player A. This is the same behavior as the normal minimax algorithm. The heuristic value is not necessarily the same as a node's return value, bestValue, due to value negation by negamax and the color parameter. The negamax node's return value is a heuristic score from the point of view of the node's current player.

Negamax scores match minimax scores for nodes where player A is about to play, and where player A is the maximizing player in the minimax equivalent. Negamax always searches for the maximum value for all its nodes. Hence for player B nodes, the minimax score is a negation of its negamax score. Player B is the minimizing player in the minimax equivalent.

Variations in negamax implementations may omit the color parameter. In this case, the heuristic evaluation function must return values from the point of view of the node's current player.

Negamax with alpha beta pruning

An animated pedagogical example showing the negamax algorithm with alpha-beta pruning. The person performing the game tree search is considered to be the one that has to move first from the current state of the game (player in this case)

Algorithm optimizations for minimax are also equally applicable for Negamax. Alpha-beta pruning can decrease the number of nodes the negamax algorithm evaluates in a search tree in a manner similar with its use with the minimax algorithm.

The pseudocode for depth-limited negamax search with alpha-beta pruning follows:[1]

01 function negamax(node, depth, α, β, color)
02     if depth = 0 or node is a terminal node
03         return color * the heuristic value of node

04     childNodes := GenerateMoves(node)
05     childNodes := OrderMoves(childNodes)
06     bestValue := −∞
07     foreach child in childNodes
08         v := −negamax(child, depth − 1, −β, −α, −color)
09         bestValue := max( bestValue, v )
10         α := max( α, v )
11         if α ≥ β
12             break
13     return bestValue
Initial call for Player A's root node
rootNegamaxValue := negamax( rootNode, depth, −∞, +∞, 1)

Alpha (α) and beta (β) represent lower and upper bounds for child node values at a given tree depth. Negamax sets the arguments α and β for the root node to the lowest and highest values possible. Other search algorithms, such as negascout and MTD-f, may initialize α and β with alternate values to further improve tree search performance.

When negamax encounters a child node value outside an alpha/beta range, the negamax search cuts off (pseudocode line 12 break statement) thereby pruning portions of the game tree from exploration. Cut offs are implicit based on the node return value, bestValue. A node value found within the range of its initial α and β is the node's exact (or true) value. This value is identical to the result the negamax base algorithm would return, without cut offs and without any α and β bounds. If a node return value is out of range, then the value represents an upper (if value ≤ α) or lower (if value ≥ β) bound for the node's exact value. Alpha-beta pruning eventually discards any value bound results. Such values do not contribute nor affect the negamax value at its root node.

This pseudocode shows the fail-soft variation of alpha-beta pruning. Fail-soft never returns α or β directly as a node value. Thus, a node value may be outside the initial α and β range bounds set with a negamax function call. In contrast, fail-hard alpha-beta pruning always limits a node value in the range of α and β.

This implementation also shows optional move ordering prior to the foreach loop that evaluates child nodes. Move ordering[2] is an optimization for alpha beta pruning that attempts to guess the most probable child nodes that yield the node's score. The algorithm searches those child nodes first. The result of good guesses is earlier and more frequent alpha/beta cut offs occur, thereby pruning additional game tree branches and remaining child nodes from the search tree.

Negamax with alpha beta pruning and transposition tables

Transposition tables selectively memorizes the values of nodes in the game tree. Transposition is a term reference that a given game board position can be reached in more than one way with differing game move sequences.

When negamax searches the game tree, and encounters the same node multiple times, a transposition table can return a previously computed value of the node, skipping potentially lengthy and duplicate re-computation of the node's value. Negamax performance improves particularly for game trees with many paths that lead to a given node in common.

The pseudo code that adds transposition table functions to negamax with alpha/beta pruning is given as follows:[1]

function negamax(node, depth, α, β, color)
    alphaOrig := α

    // Transposition Table Lookup; node is the lookup key for ttEntry
    ttEntry := TranspositionTableLookup( node )
    if ttEntry is valid and ttEntry.depth ≥ depth
        if ttEntry.Flag = EXACT
            return ttEntry.Value
        else if ttEntry.Flag = LOWERBOUND
            α := max( α, ttEntry.Value)
        else if ttEntry.Flag = UPPERBOUND
            β := min( β, ttEntry.Value)
        endif
        if α ≥ β
            return ttEntry.Value
    endif

    if depth = 0 or node is a terminal node
        return color * the heuristic value of node

    bestValue := -∞
    childNodes := GenerateMoves(node)
    childNodes := OrderMoves(childNodes)
    foreach child in childNodes
        v := -negamax(child, depth - 1, -β, -α, -color)
        bestValue := max( bestValue, v )
        α := max( α, v )
        if α ≥ β
            break

    // Transposition Table Store; node is the lookup key for ttEntry
    ttEntry.Value := bestValue
    if bestValue ≤ alphaOrig
        ttEntry.Flag := UPPERBOUND
    else if bestValue ≥ β
        ttEntry.Flag := LOWERBOUND
    else
        ttEntry.Flag := EXACT
    endif
    ttEntry.depth := depth	
    TranspositionTableStore( node, ttEntry )

    return bestValue
Initial call for Player A's root node
rootNegamaxValue := negamax( rootNode, depth, −∞, +∞, 1)

Alpha/beta pruning and maximum search depth constraints in negamax can result in partial, inexact, and entirely skipped evaluation of nodes in a game tree. This complicates adding transposition table optimizations for negamax. It's insufficient to track only the node's bestValue in the table, because bestValue may not be the node's true value. The code therefore must preserve and restore the relationship of bestValue with alpha/beta parameters and the search depth for each transposition table entry.

Transposition tables are typically lossy and will omit or overwrite previous values of certain game tree nodes in its tables. This is necessary since the number of nodes negamax visits often far exceeds the transposition table size. Lost or omitted table entries are non critical and won't affect the negamax result. However, lost entries may require negamax to re-compute certain game tree node values more frequently, thus affecting performance.

References

  1. 1 2 3 Breuker, Dennis M. Memory versus Search in Games, Maastricht University, October 16, 1998
  2. Schaeffer, Jonathan The History Heuristic and Alpha-Beta Search Enhancements in Practice, IEEE Transactions on Pattern Analysis and Machine Intelligence, 1989

External links

This article is issued from Wikipedia - version of the 11/13/2016. The text is available under the Creative Commons Attribution/Share Alike but additional terms may apply for the media files.