def minimax(self, side, board, depth, max_depth):
mov_gen = shogi.Rule().gen_moves
ev = self.eval
if depth == max_depth:
result = self.lookup.get(board.__hash__(), None)
if result is None:
self.lookup[board.__hash__()] = ev(board)
return self.lookup[board.__hash__()]
else:
if side == 1:
best = -1000
moves = mov_gen(side, board)
if len(moves) == 0:
return -1000
for b in moves:
now = self.minimax(-side, b, depth + 1, max_depth)
if now > best:
best = now
if depth == 0:
self.result = b
return best
else:
best = 1000
moves = mov_gen(side, board)
if len(moves) == 0:
return 1000
for b in moves:
now = self.minimax(-side, b, depth + 1, max_depth)
if now < best:
best = now
if depth == 0:
self.result = b
return best
def make_move(self, side, board):
mov_gen = shogi.Rule().gen_moves
moves = mov_gen(side, board)
if len(moves) == 0:
return None
q = []
for b in moves:
q.append(model.predict(b.vectorize().reshape((1, 13 * 11))))
if side == 1:
return moves[q.index(max(q))]
if side == -1:
return moves[q.index(min(q))]
hidden_size = 512
batch_size = 50
model = Sequential()
model.add(Dense(hidden_size, input_shape=(13 * 11, ), activation='relu'))
model.add(Dense(hidden_size, activation='relu'))
model.add(Dense(hidden_size, activation='relu'))
model.add(Dense(hidden_size, activation='relu'))
model.add(Dense(1))
model.compile(Adam(lr=.2), "mse")
# If you want to continue training from a previous model, just uncomment the line bellow
# model.load_weights("manmodel.h5")
# Define environment/game
r = shogi.Rule()
# Initialize experience replay object
exp_replay = ExperienceReplay(max_memory=max_memory)
# Train
win_cnt = 0
for e in range(epoch):
loss = 0.
board = shogi.Board([[-2, -1, -3], [0, -4, 0], [0, 4, 0], [3, 1, 2]],
[])
game_over = False
# get initial input
input_t = board
while not game_over:
def make_move(self, side, board):
randint = random.randint
mov_gen = shogi.Rule().gen_moves
l = mov_gen(side, board)
return None if len(l) == 0 else l[randint(0, len(l) - 1)]