def choose_action(self, state, model):
rbs = rbg_game.resettable_bitarray_stack()
moves = state.get_all_moves(rbs)
boards = []
for mv in moves:
boardcp = state.copy()
boardcp.apply_with_keeper(mv, rbs)
boards.append(model_utilis.parse_game_state(boardcp))
values = model(torch.cat(boards, 0).to(self.device))
max_id = torch.argmax(values)
return moves[max_id], values[max_id]
def test(test_config):
rbs = rbg_game.resettable_bitarray_stack()
model_1 = test_config['model_1']
player_1 = test_config['player_1']
model_2 = test_config['model_2']
player_2 = test_config['player_2']
num_games = test_config['num_games']
player_1_number = 1
player_2_number = 2
wins = 0
draws = 0
for game_number in tqdm(range(num_games)):
game_state = rbg_game.new_game_state()
while True:
if game_state.get_current_player() == 0:
break
if game_state.get_current_player() == player_1_number:
move_rbs, _ = player_1.choose_action(game_state, model_1)
make_move(game_state, move_rbs, rbs)
else:
move_rbs, _ = player_2.choose_action(game_state, model_2)
make_move(game_state, move_rbs, rbs)
if game_state.get_player_score(player_1_number) == 100:
wins += 1
if game_state.get_player_score(player_1_number) == 50:
draws += 1
player_2_number, player_1_number = player_1_number, player_2_number
print('Wins=', wins)
print('Draws=', draws)
print('Wins+Draws=', wins + draws)
print('Win+Draw rate=', (wins + draws) / num_games)
def train(train_config):
rbs = rbg_game.resettable_bitarray_stack()
model_1 = train_config['model_1']
trainer_1 = train_config['trainer_1']
player_1 = train_config['player_1']
model_2 = train_config['model_2']
trainer_2 = train_config['trainer_2']
player_2 = train_config['player_2']
model_name = train_config['model_name']
num_games = train_config['num_games']
save_model_every_n_iterations = train_config[
'save_model_every_n_iterations']
save_path = train_config['save_path']
player_1_number = 1
player_2_number = 2
for game_number in tqdm(range(num_games)):
game_state = rbg_game.new_game_state()
while True:
if game_state.get_current_player() == 0:
break
old_state = game_state.copy()
if game_state.get_current_player() == player_1_number:
move_rbs, move_value = player_1.choose_action(
game_state, model_1)
make_move(game_state, move_rbs, rbs)
if game_state.get_current_player() == 0:
trainer_1.add_to_history(
replay.MovePerformed(
old_state, move_rbs, game_state.copy(), move_value,
game_state.get_player_score(player_1_number) /
100))
else:
trainer_1.add_to_history(
replay.MovePerformed(old_state, move_rbs,
game_state.copy(), move_value))
else:
move_rbs, move_value = player_2.choose_action(
game_state, model_2)
make_move(game_state, move_rbs, rbs)
if game_state.get_current_player() == 0:
trainer_2.add_to_history(
replay.MovePerformed(
old_state, move_rbs, game_state.copy(), move_value,
game_state.get_player_score(player_2_number) /
100))
else:
trainer_2.add_to_history(
replay.MovePerformed(old_state, move_rbs,
game_state.copy(), move_value))
trainer_1.learn(game_state.get_player_score(player_1_number) / 100)
trainer_2.learn(game_state.get_player_score(player_2_number) / 100)
player_2_number, player_1_number = player_1_number, player_2_number
if game_number % save_model_every_n_iterations == 0:
save_model(save_path, model_name + '_model_1', model_1)
save_model(save_path, model_name + '_model_2', model_2)
import rbg_game
import random
import time
bitarraystack = rbg_game.resettable_bitarray_stack()
# Visits all the game tree of a given game to the given depth
# Returns a pair of numbers of leaves visited and number of nodes visited
def perft(state, depth):
if not state:
return [0, 0]
if depth == 0:
return [1, 1]
moves = state.get_all_moves(bitarraystack)
result = [0, 0]
result[1] += 1
for move in moves:
next_state = state.copy()
next_state.apply_with_keeper(move, bitarraystack)
rec_result = perft(next_state, depth - 1)
result[0] += rec_result[0]
result[1] += rec_result[1]
return result
# Play the game till the end from the current state.
# This modifies the given state.
def playout(state):
nodes = 0
while True:
def choose_action(self, game_state):
rbs = rbg_game.resettable_bitarray_stack()
move = random.choice(game_state.get_all_moves(rbs))
return (move, None)