def nnet_learning(self):
memory_size = 2000
self.network.eval()
for e in range(self.epochs):
print("Epoch : " + str(e))
self.num_sim = self.num_sim + e * 5
if self.num_sim > 300:
self.num_sim = 300
for s in range(self.num_selfplay):
print(" Game number : " + str(s))
self.training_examples_all.extend(self.selfplay())
self.save_examples(self.example_num)
self.example_num += 1
self.training_examples_all = []
self.mcts = mcts(self.c4_utils, self.network, 1.5,
self.num_sim)
self.c4_utils.turn = 0
print(" End of self play")
self.network.train()
self.network.saveweights(filename='wt_temp.pth.tar')
self.network_old.loadweights(filename='wt_temp.pth.tar')
net_old_mcts = mcts(self.c4_utils, self.network_old, 1.5,
self.num_sim)
training_examples = self.load_examples(self.example_num)
self.network.train_net(training_examples)
training_examples = []
net_mcts = mcts(self.c4_utils, self.network, 1.5, self.num_sim)
self.network.eval()
self.network_old.eval()
c = Compete(net_old_mcts, net_mcts, self.c4_utils)
print(" Competition")
net_old_win, net_win, draw = c.play(self.num_competitions)
print(
" New network wins : %d, Old network wins : %d, Draw : %d "
% (net_win, net_old_win, draw))
if (net_old_win > net_win):
print(" Rejecting new model")
self.network.loadweights(filename='wt_temp.pth.tar')
else:
self.network.saveweights(filename='wt_best.pth.tar')
def self_play(model: keras.layers.Layer, config: Config) -> Game:
game = Game()
while not game.terminal():
action, root = mcts(game, model, config)
game.apply(action)
game.store_search_statistics(root)
return game
def __init__(self, network, num_selfplay, epochs, chess_utils,
num_competitions, num_sim_):
self.num_sim = num_sim_
self.network = network
self.chess_utils = chess_utils
self.network_old = self.network.__class__(epochs)
self.mcts = mcts(self.chess_utils, self.network, 1.5, self.num_sim)
self.num_selfplay = num_selfplay
self.epochs = epochs
self.num_competitions = num_competitions
self.training_examples_all = []
self.update_threshold = 0.5
self.example_num = 0
def run_once(debug=True):
# Randomly assigns actors, places, and items for story
root_state = random_state(4, 4)
# Initialize Root Node - Possible Methods boolean MUST BE TRUE
root_node = TreeNode(root_state, parent_edge=None, possible_methods=True)
# Total methods in story
num_methods = len(root_node.possible_methods)
"""
The following
max_numsim = max_expansion * thres
max_iter : Number of sentances in story = number of story nodes - 1 = number of story edges
max_expansion : Number of expansions in search
max_simlength : Maximum length of rollout
C : Exploration Constant for selection
thres : Minimum MCTS Visits for node expansion
"""
# Perform Monte Carlo - returns final node and whole story
max_expansion = 250
if max_expansion < len(root_node.possible_methods):
raise ValueError(
"Max exp ({}) should be greater than num methods({})".format(
max_expansion, len(root_node.possible_methods)))
max_iter = 15
max_simlength = 20
C = 1
thres = 40
minlambda = 0.95
s = Story(root_node)
print(s.create_expository())
#print("Max iteration: {}\nMax Expansion: {}\nMax simulation length: {}\nC: {}\nThreshold: {}".format(max_iter, max_expansion, max_simlength, C, thres))
n, s = mcts(root_node,
max_iter,
max_expansion,
max_simlength,
C,
thres,
mixlambda,
debug=False)
# Print out results
#if debug:
# print(s)
# print(n.believability)
# print(n.value)
# print(percent_goals_satisfied(n, GOALS))
return (n, s)
def __init__(self, c4_utils, network, num_sim, player_x):
self.round = 1
self.finished = False
self.winner = None
self.mcts = mcts(c4_utils, network, 1.5, num_sim)
name_mmx = "Minimax"
name_mc = "Monte-Carlo Sim - RL"
if player_x == "m":
self.players[0] = Player(name_mc, self.colors[0], self.mcts)
diff = int(input("Enter difficulty for the Minimax AI (1 - 4) "))
self.players[1] = AIPlayer(name_mmx, self.colors[1], diff + 1)
else:
diff = int(input("Enter difficulty for the Minimax AI (1 - 4) "))
self.players[0] = AIPlayer(name_mmx, self.colors[0], diff + 1)
self.players[1] = Player(name_mc, self.colors[1], self.mcts)
# # do cross-platform clear screen
# os.system( [ 'clear', 'cls' ][ os.name == 'nt' ] )
# print(u"Welcome to {0}!".format(self.game_name))
# print("Should Player 1 be a Minimax(Benchmark) or Monte-Carlo(Our program)?")
# while self.players[0] == None:
# choice = str(input("Type 'MINIMAX (x)' or 'MONTE-CARLO (m)': "))
# if choice == "M" or choice.lower() == "m":
# diff = int(input("Enter difficulty for the Minimax AI (1 - 4) "))
# self.players[0] = Player(name_mc, self.colors[0],self.mcts)
# self.players[1] = AIPlayer(name_mmx, self.colors[1], diff +1 )
# elif choice == "X" or choice.lower() == "x":
# diff = int(input("Enter difficulty for the Minimax AI (1 - 4) "))
# self.players[0] = AIPlayer(name_mmx, self.colors[0], diff+1)
# self.players[1] = Player(name_mc ,self.colors[1], self.mcts)
# else:
# print("Invalid choice, please try again")
print("{0} will be {1}".format(self.players[0].name, self.colors[0]))
# x always goes first (arbitrary choice on my part)
self.turn = self.players[0]
self.board = []
for i in range(6):
self.board.append([])
for j in range(7):
self.board[i].append(' ')