def take_action(self, env, a=None):
s = env.current_state
as_pairs = env.get_as_pairs()
e = np.random.rand()
if e < self.policy.eps:
rand = np.random.randint(len(as_pairs))
a = as_pairs[rand][0]
sn = as_pairs[rand][1]
_, v = search.alphabeta_batch_hist(self.V, faster_featurize,
Environment(sn), list(env.hist),
self.depth - 1, -float('inf'),
float('inf'))
else:
a, v = self.get_av_pairs(env)
r, s_n = env.perform_action(a)
# r_w reward from whites perspective
r_w = (-1 if env.get_turn() else 1) * r # invert value if white
#print r_w
#print r
self.data_thread.append((s, r_w, v))
if Environment.terminal_state(s_n):
#print 'Reward for white: ',r_w
self.data_thread.append((s_n, r_w, r_w))
self.data_thread.set_update()
if env.result() == '1-0' or env.result() == '0-1':
#env.draw()
self.data_thread.set_win()
return s, a, r, s_n
def take_action(self, env, a=None):
s = env.current_state
as_pairs = env.get_as_pairs()
e = np.random.rand()
if e < self.policy.eps:
rand = np.random.randint(len(as_pairs))
a = as_pairs[rand][0]
sn = as_pairs[rand][1]
_, v = search.alphabeta_batch_hist(self.V, faster_featurize,
Environment(sn), list(env.hist),
self.depth - 1, -float('inf'),
float('inf'))
else:
a, v = self.get_av_pairs(env)
if a is None:
for st in env.hist:
print chess.Board.from_epd(st)
print('Value: {}, random e: {}, epsilon: {}').format(
v, e, self.policy.eps)
print 'as pairs: {}'.format(as_pairs)
env.draw()
r, s_n = env.perform_action(a)
# r_w reward from whites perspective
r_w = (-1 if env.get_turn() else 1) * r # invert value if white
self.data_thread.append((s, r, v))
if Environment.terminal_state(s_n):
self.data_thread.append((s_n, r, r))
self.data_thread.set_outcome(env.outcome())
return s, a, r, s_n
def get_av_pairs(self, env):
ao, o = search.alphabeta_outcome(None, None, env.current_state,
self.depth, -float('inf'),
float('inf'))
a, v = search.alphabeta_batch_hist(self.approx.value, faster_featurize,
env, list(env.hist),
settings.params['OC_DEPTH'],
-float('inf'), float('inf'))
if o > 0:
a = ao
return [(a, v)]
def take_action(self, env, a=None):
ao, o = search.alphabeta_outcome(None, None, env.current_state,
self.depth, -float('inf'),
float('inf'))
if o > 0:
env.perform_action(ao)
else:
a, v = search.alphabeta_batch_hist(self.approx.value,
faster_featurize, env,
list(env.hist), self.depth,
-float('inf'), float('inf'))
env.perform_action(a)
def get_av_pairs(self, env):
t = time.time()
a, v = search.alphabeta_batch_hist(self.V, faster_featurize, env,
list(env.hist), self.depth,
-float('inf'), float('inf'))
t2 = time.time()
ao, o = search.alphabeta_outcome(None, None, env.current_state,
settings.params['OC_DEPTH'],
-float('inf'), float('inf'))
t3 = time.time()
if o > 0:
a = ao
env.draw()
print a, o, (t3 - t2) / (t2 - t)
return a, v
def get_av_pairs(self, env):
return search.alphabeta_batch_hist(self.V, faster_featurize, env,
list(env.hist), self.depth,
-float('inf'), float('inf'))
def take_action(self, env, a=None):
a, v = search.alphabeta_batch_hist(self.approx.value, faster_featurize,
env, list(env.hist), self.depth,
-float('inf'), float('inf'))
env.perform_action(a)