def __init__(self, *args, **kwargs):
self.epsilon = 0.2
if "epsilon" in kwargs:
self.epsilon = kwargs["epsilon"] or self.epsilon
del kwargs["epsilon"]
self.multiplier = 3
if "multiplier" in kwargs:
self.multiplier = kwargs["multiplier"] or self.multiplier
del kwargs["multiplier"]
DualAgent.__init__(self, *args, **kwargs)
self.tree = MonteCarloTree(self.sess, self.epsilon, self.multiplier)
def get_random_action(state):
y_p = DualAgent.get_dist(agent, state)
c = np.random.choice(225, p=y_p)
x, y = np.unravel_index(c, dims=(15, 15))
assert state.board[x, y] == 0, "total prob %f that prob %f" % (y_p.sum(),
y_p[c])
return x, y
from agent import Agent
from scipy.io import loadmat
from dual_agent import DualAgent
from dual_network import export_meta
import numpy as np
import tensorflow as tf
import data_util as util
""" begin training """
with tf.Session() as sess:
export_meta("dualsup")
agent = DualAgent(sess, "dualsup")
matlab = loadmat("dual_minimax")
print("processing data")
data = util.TenaryData(matlab["X"], matlab["Y"], matlab["V"][0])
del matlab
print("processing complete")
for i in range(4001):
x_b, y_b, v_b = data.next_batch(256)
if i % 10 == 0:
x_b, y_b, v_b = data.test_batch(1024)
pl, vl, rl, l = agent.loss(x_b, y_b, v_b)
print("\nstep %d analysis" % i)
print(">>> policy loss: %f" % pl)
print(">>> value loss: %f" % vl)
print(">>> regularization loss: %f" % rl)
print(">>> loss: %f" % l)
print(">>> accuracy: %f" % agent.accuracy(x_b, y_b))
parser.add_argument("model_name_1", type=str)
parser.add_argument("model_name_2", type=str)
parser.add_argument("--epsilon", "-e", type=float)
parser.add_argument("--chkpnt1", "-c1", type=int)
parser.add_argument("--chkpnt2", "-c2", type=int)
parser.add_argument("--num_games", "-n", default=100, type=int)
parser.add_argument('--save', '-s', action='store_true')
args = parser.parse_args()
a_graph = tf.Graph()
b_graph = tf.Graph()
a_sess = tf.Session(graph=a_graph)
b_sess = tf.Session(graph=b_graph)
with a_sess.as_default():
with a_graph.as_default():
a_agent = DualAgent(a_sess, args.model_name_1, chkpnt=args.chkpnt1)
with b_sess.as_default():
with b_graph.as_default():
b_agent = MCTSAgent(b_sess,
args.model_name_2,
chkpnt=args.chkpnt2,
epsilon=args.epsilon)
print("ARENA: DUAL %s-%d VERSES MCTS %s-%d" %
(a_agent.model_name, a_agent.chkpnt, b_agent.model_name, b_agent.chkpnt))
stat = np.zeros(shape=(2, 2), dtype=np.int)
for i in range(args.num_games):
t = time()
s = State()
a_is_black = (i % 2 == 0)
while not s.end and len(s.history) < 225:
from dual_agent import DualAgent
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("id", type=int)
args = parser.parse_args()
ID = args.id
NUM_GAMES_GEN = 200
NUM_SIM = 100
X = []
V = []
with tf.Session() as sess:
agent = DualAgent(sess, "treesup")
for i in range(NUM_GAMES_GEN):
print("sample %d" % i)
s = State()
while not s.end and len(s.history) < 225:
s.move(*agent.get_safe_action(s))
assert not s.violate, "we have a violated game"
which = np.random.randint(len(s.history))
h = s.history
s = State()
for k in range(which):
s.move(*h[k])
score = 0
for j in range(NUM_SIM):
t = s.copy()
while not t.end and len(t.history) < 225:
parser.add_argument("model_name_2", type=str)
parser.add_argument("--max_width", "-w", type=int)
parser.add_argument("--max_depth", "-d", type=int)
parser.add_argument("--chkpnt1", "-c1", type=int)
parser.add_argument("--chkpnt2", "-c2", type=int)
parser.add_argument("--num_games", "-n", default=100, type=int)
parser.add_argument('--save', '-s', action='store_true')
args = parser.parse_args()
a_graph = tf.Graph()
b_graph = tf.Graph()
a_sess = tf.Session(graph=a_graph)
b_sess = tf.Session(graph=b_graph)
with a_sess.as_default():
with a_graph.as_default():
a_agent = DualAgent(a_sess, args.model_name_1, chkpnt=args.chkpnt1)
with b_sess.as_default():
with b_graph.as_default():
b_agent = MinimaxNetworkAgent(b_sess, args.model_name_2, chkpnt=args.chkpnt2, max_width=args.max_width, max_depth=args.max_depth)
print("ARENA: DUAL %s-%d VERSES MNA %s-%d" % (a_agent.model_name, a_agent.chkpnt, b_agent.model_name, b_agent.chkpnt))
stat = np.zeros(shape=(2, 2), dtype=np.int)
for i in range(args.num_games):
t = time()
s = State()
a_is_black = (i % 2 == 0)
while not s.end and len(s.history) < 225:
if a_is_black == (s.player > 0):
with a_sess.as_default():
s.move(*a_agent.get_safe_action(s))
else:
parser = argparse.ArgumentParser()
parser.add_argument("model_name_1", type=str)
parser.add_argument("model_name_2", type=str)
parser.add_argument("--chkpnt1", "-c1", type=int)
parser.add_argument("--chkpnt2", "-c2", type=int)
parser.add_argument("--num_games", "-n", default=100, type=int)
parser.add_argument('--save', '-s', action='store_true')
args = parser.parse_args()
a_graph = tf.Graph()
b_graph = tf.Graph()
a_sess = tf.Session(graph=a_graph)
b_sess = tf.Session(graph=b_graph)
with a_sess.as_default():
with a_graph.as_default():
a_agent = DualAgent(a_sess, args.model_name_1, chkpnt=args.chkpnt1)
with b_sess.as_default():
with b_graph.as_default():
b_agent = DualAgent(b_sess, args.model_name_2, chkpnt=args.chkpnt2)
print("ARENA: %s-%d VERSES %s-%d" % (a_agent.model_name, a_agent.chkpnt, b_agent.model_name, b_agent.chkpnt))
stat = np.zeros(shape=(2, 2), dtype=np.int)
for i in range(args.num_games):
t = time()
s = State()
a_is_black = (i % 2 == 0)
while not s.end and len(s.history) < 225:
if a_is_black == (s.player > 0):
with a_sess.as_default():
s.move(*a_agent.get_action(s))
else: