mirror of
https://github.com/luau-lang/luau.git
synced 2024-11-15 14:25:44 +08:00
228 lines
9.9 KiB
Lua
228 lines
9.9 KiB
Lua
|
local bench = script and require(script.Parent.bench_support) or require("bench_support")
|
||
|
|
|
||
|
|
function test()
|
||
|
|
|
||
|
|
-- https://github.com/stefandd/Tic4
|
||
|
|
local negaMax = {maxdepth = 4, minsearchpos = 0, numsearchpos = 0}
|
||
|
|
negaMax.__index = negaMax
|
||
|
|
|
||
|
|
function negaMax:evaluate(board, depth)
|
||
|
|
--[[
|
||
|
|
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 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.
|
||
|
|
--]]
|
||
|
|
print ("This function needs to be implemented!")
|
||
|
|
end
|
||
|
|
|
||
|
|
function negaMax:move_candidates(board, side_to_move)
|
||
|
|
print ("This function needs to be implemented!")
|
||
|
|
end
|
||
|
|
|
||
|
|
function negaMax:make_move(board, side_to_move, move)
|
||
|
|
print ("This function needs to be implemented!")
|
||
|
|
end
|
||
|
|
|
||
|
|
function negaMax:negaMax(board, side_to_move, depth, alpha, beta) -- side_to_move: e.g. 1 is blue, -1 is read
|
||
|
|
--
|
||
|
|
-- init vars for root call
|
||
|
|
--
|
||
|
|
if not depth then -- root call
|
||
|
|
depth = 0
|
||
|
|
alpha = -math.huge
|
||
|
|
beta = math.huge
|
||
|
|
self.numsearchpos = 0 -- reset call counter
|
||
|
|
end
|
||
|
|
--
|
||
|
|
-- test if the node is terminal (i.e. full board or win)
|
||
|
|
--
|
||
|
|
local best_move = -1
|
||
|
|
local score, is_term_node = self:evaluate(board, depth)
|
||
|
|
-- we abort the recursion if this is a terminal node, or if one of the search abort conditions are met
|
||
|
|
--
|
||
|
|
if is_term_node or depth == self.maxdepth then
|
||
|
|
return side_to_move*score, best_move, is_term_node
|
||
|
|
end
|
||
|
|
--
|
||
|
|
-- if not terminal node, eval child nodes
|
||
|
|
--
|
||
|
|
local moves = self:move_candidates(board, side_to_move)
|
||
|
|
score = -math.huge
|
||
|
|
|
||
|
|
for _, analyzed_move in pairs(moves) do -- iterate over all boards
|
||
|
|
self.numsearchpos = self.numsearchpos + 1
|
||
|
|
local b = self:make_move(board, side_to_move, analyzed_move)
|
||
|
|
local move_score, _, _ = -self:negaMax(b, -side_to_move, depth+1, -beta, -alpha)
|
||
|
|
if move_score > score then
|
||
|
|
score = move_score
|
||
|
|
best_move = analyzed_move
|
||
|
|
end
|
||
|
|
-- disable alpha-beta pruning
|
||
|
|
--
|
||
|
|
alpha = math.max(alpha, score)
|
||
|
|
if alpha >= beta then
|
||
|
|
break
|
||
|
|
end
|
||
|
|
--
|
||
|
|
end
|
||
|
|
if depth == 0 then
|
||
|
|
--
|
||
|
|
-- exit for root call (depth == 0)
|
||
|
|
--
|
||
|
|
-- debug stuff
|
||
|
|
--print(string.format("---- Negamax: node info, depth: %d, side: %d, score: %d, best move: %d", depth, side_to_move, score, best_move))
|
||
|
|
--print_board(board)
|
||
|
|
--print(string.format("----"))
|
||
|
|
print("Analyzed positions: " .. self.numsearchpos)
|
||
|
|
end
|
||
|
|
return score, best_move, game_over
|
||
|
|
end
|
||
|
|
|
||
|
|
local empty_board = {0,0,0,0,
|
||
|
|
0,0,0,0,
|
||
|
|
0,0,0,0,
|
||
|
|
0,0,0,0} -- 16 empty positions
|
||
|
|
|
||
|
|
----------- helper methods
|
||
|
|
|
||
|
|
function copy_board(board)
|
||
|
|
local copy = {}
|
||
|
|
for i = 1, #board do
|
||
|
|
copy[i] = board[i]
|
||
|
|
end
|
||
|
|
return copy
|
||
|
|
end
|
||
|
|
|
||
|
|
function print_board(board)
|
||
|
|
gboard = {}
|
||
|
|
for i = 1, #board do
|
||
|
|
if board[i] == 0 then gboard[i] = '.'
|
||
|
|
elseif board[i] == 1 then gboard[i] = 'x'
|
||
|
|
else gboard[i] = 'o'
|
||
|
|
end
|
||
|
|
end
|
||
|
|
print(string.format("\n%s %s %s %s\n%s %s %s %s\n%s %s %s %s\n%s %s %s %s\n", unpack(gboard)))
|
||
|
|
end
|
||
|
|
|
||
|
|
function is_board_full(board)
|
||
|
|
for i = 1, #board do
|
||
|
|
if board[i] == 0 then
|
||
|
|
return false
|
||
|
|
end
|
||
|
|
end
|
||
|
|
return true
|
||
|
|
end
|
||
|
|
|
||
|
|
----------- implement negaMax methods
|
||
|
|
|
||
|
|
negaMax.index_quadruplets = {
|
||
|
|
{1,2,3,4}, {5,6,7,8}, {9,10,11,12}, -- rows
|
||
|
|
{13,14,15,16}, {1,5,9,13}, {2,6,10,14}, -- cols
|
||
|
|
{3,7,11,15}, {4,8,12,16}, {1,6,11,16}, {4,7,10,13}, -- diags
|
||
|
|
{1,2,5,6}, {2,3,6,7}, {3,4,7,8}, -- squares
|
||
|
|
{5,6,9,10}, {6,7,10,11}, {7,8,11,12},
|
||
|
|
{9,10,13,14}, {10,11,14,15}, {11,12,15,16}
|
||
|
|
}
|
||
|
|
|
||
|
|
function negaMax:evaluate(board, depth) -- return format is score, is_terminal_position
|
||
|
|
--[[
|
||
|
|
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 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.
|
||
|
|
--]]
|
||
|
|
local player_plus_score, player_minus_score = 0, 0
|
||
|
|
local game_won = false
|
||
|
|
for _, curr_qdr in pairs(negaMax.index_quadruplets) do -- iterate over all index quadruplets
|
||
|
|
-- count the empty positions and positions occupied by the side whos move it is
|
||
|
|
local player_plus_fields, player_minus_fields, empties = 0, 0, 0
|
||
|
|
for _, index in pairs(curr_qdr) do -- iterate over all indices
|
||
|
|
if board[index] == 0 then
|
||
|
|
empties = empties + 1
|
||
|
|
elseif board[index] == 1 then
|
||
|
|
player_plus_fields = player_plus_fields + 1
|
||
|
|
elseif board[index] == -1 then
|
||
|
|
player_minus_fields = player_minus_fields + 1
|
||
|
|
end
|
||
|
|
end
|
||
|
|
-- evaluate the quadruplets score by looking at empty vs occupied positions
|
||
|
|
if empties == 3 then
|
||
|
|
if player_plus_fields == 1 then
|
||
|
|
player_plus_score = player_plus_score + 3
|
||
|
|
elseif player_minus_fields == 1 then
|
||
|
|
player_minus_score = player_minus_score + 3
|
||
|
|
end
|
||
|
|
elseif empties == 2 then
|
||
|
|
if player_plus_fields == 2 then
|
||
|
|
player_plus_score = player_plus_score + 13
|
||
|
|
elseif player_minus_fields == 2 then
|
||
|
|
player_minus_score = player_minus_score + 13
|
||
|
|
end
|
||
|
|
elseif empties == 1 then
|
||
|
|
if player_plus_fields == 3 then
|
||
|
|
player_plus_score = player_plus_score + 31
|
||
|
|
elseif player_minus_fields == 3 then
|
||
|
|
player_minus_score = player_minus_score + 31
|
||
|
|
end
|
||
|
|
elseif empties == 0 then
|
||
|
|
-- check for winning situations
|
||
|
|
if player_plus_fields == 4 then
|
||
|
|
player_plus_score = 999-depth
|
||
|
|
player_minus_score = 0
|
||
|
|
game_won = true
|
||
|
|
break
|
||
|
|
elseif player_minus_fields == 4 then
|
||
|
|
-- this should not happen if there is a proper terminal node detection!
|
||
|
|
player_plus_score = 0
|
||
|
|
player_minus_score = 999-depth
|
||
|
|
game_won = true
|
||
|
|
break
|
||
|
|
end
|
||
|
|
end
|
||
|
|
end
|
||
|
|
-- return format is score, is_terminal_position
|
||
|
|
if not game_won and is_board_full(board) then
|
||
|
|
return 0, true -- DRAW
|
||
|
|
else
|
||
|
|
return (player_plus_score - player_minus_score), game_won -- >0 is good for player 1 [+], <0 means good for the other player (player 2 [-]))
|
||
|
|
end
|
||
|
|
end
|
||
|
|
|
||
|
|
function negaMax:move_candidates(board, side_to_move)
|
||
|
|
local moves = {}
|
||
|
|
for i = 1, #board do
|
||
|
|
if board[i] == 0 then -- empty?
|
||
|
|
moves[#moves + 1] = i -- save move that was made
|
||
|
|
end
|
||
|
|
end
|
||
|
|
return moves
|
||
|
|
end
|
||
|
|
|
||
|
|
function negaMax:make_move(board, side_to_move, move)
|
||
|
|
local copy = copy_board(board)
|
||
|
|
copy[move] = side_to_move
|
||
|
|
return copy
|
||
|
|
end
|
||
|
|
|
||
|
|
local human_player = 1
|
||
|
|
local AI_player = -human_player
|
||
|
|
local game_board = copy_board(empty_board)
|
||
|
|
local curr_move = -1
|
||
|
|
local curr_player = human_player -- human player goes first
|
||
|
|
local score = 0
|
||
|
|
local stop_loop = false
|
||
|
|
local game_over = false
|
||
|
|
|
||
|
|
negaMax.maxdepth = 5
|
||
|
|
|
||
|
|
local t0 = os.clock()
|
||
|
|
score, curr_move = negaMax:negaMax(game_board, curr_player)
|
||
|
|
local t1 = os.clock()
|
||
|
|
|
||
|
|
return t1-t0
|
||
|
|
end
|
||
|
|
|
||
|
|
bench.runCode(test, "tictactoe")
|