Conversion of minimax with alpha beta pruning to negamax

Question

I've written a minimax algorithm with alpha beta pruning for the game Checkers, and now I'm trying to rewrite it using the negamax approach. I'm expecting the two to be equivalent, since negamax is just a technique to write the minimax. But for some reason my two algorithms behave differently. When I run them both on the same input, the negamax version seems to evaluate more states, so I think something must be wrong with the alpha beta pruning.

The code below shows both algorithms (minimax and negamax functions), and at the bottom the play function from which I call them. The evaluate function is the basic heuristic which I use to evaluate states in both algorithms.

Any help with spotting the error would be much appriciated.

#include "player.hpp"
#include <algorithm>
#include <limits>
#include <cstdlib>

namespace checkers
{

int evaluatedStates = 0;

int evaluate(const GameState &state)
{
    // FIXME: Improve heuristics.
    int redScore = 0;
    int whiteScore = 0;
    int piece = 0;
    for (int i = 1; i <= 32; ++i)
    {
        piece = state.at(i);
        if (piece & CELL_RED) {
            ++redScore;
            if (piece & CELL_KING)
                redScore += 2;   // King bonus.
        } else if (piece & CELL_WHITE) {
            ++whiteScore;
            if (piece & CELL_KING)
                whiteScore += 2; // King bonus.
        }
    }
    return state.getNextPlayer() == CELL_RED ? whiteScore - redScore : redScore - whiteScore;
}

int minimax(const GameState &state, int depth, int a, int b, bool max)
{
    if (depth == 0 || state.isEOG()) {
        ++evaluatedStates;
        return evaluate(state);
    }
    std::vector<GameState> possibleMoves;
    state.findPossibleMoves(possibleMoves);
    if (max) {
        for (const GameState &move : possibleMoves) {
            a = std::max(a, minimax(move, depth - 1, a, b, false));
            if (b <= a)
                return b; // β cutoff.
        }
        return a;
    } else {
        for (const GameState &move : possibleMoves) {
            b = std::min(b, minimax(move, depth - 1, a, b, true));
            if (b <= a)
                return a; // α cutoff.
        }
        return b;
    }
}

int negamax(const GameState &state, int depth, int a, int b)
{
    if (depth == 0 || state.isEOG()) {
        ++evaluatedStates;
        return evaluate(state);
    }
    std::vector<GameState> possibleMoves;
    state.findPossibleMoves(possibleMoves);
    for (const GameState &move : possibleMoves) {
        a = std::max(a, -negamax(move, depth - 1, -b, -a));
        if (b <= a)
            return b; // β cutoff.
    }
    return a;
}

GameState Player::play(const GameState &pState, const Deadline &pDue)
{
    GameState bestMove(pState, Move());

    std::vector<GameState> possibleMoves;
    pState.findPossibleMoves(possibleMoves);

    int a = -std::numeric_limits<int>::max();
    int b = std::numeric_limits<int>::max();
    for (const GameState &move : possibleMoves) {
        int v = negamax(move, 10, a, b);
        //int v = minimax(move, 10, a, b, false);
        if (v > a) {
            a = v;
            bestMove = move;
        }
    }
    std::cerr << "Evaluated states: " << evaluatedStates << std::endl;
    return bestMove;
}

/*namespace checkers*/ }

Answer 1

Your minimax() and negamax() functions are correct. I assume that state.getNextPlayer() returns the player who has to move next. That means that your evaluate() and negamax() functions return a score from the perspective of that player.

However, the minimax() returns a score from the perspective of max. So if you try uncommenting minimax() in your play() function, that would lead to a bug

//int v = negamax(move, 10, a, b);
int v = minimax(move, 10, a, b, false); // assumes perspective of min player
                                ^^^^^

if (v > a) {                            // assumes perspective of max player
    a = v;
    bestMove = move;
}

Replacing the call to minimax() with a true parameter should solve it:

int v = minimax(move, 10, a, b, true); // assumes perspective of max player