def run_sim(turn): # simulate game forward
arch.sess.run(arch.session_backup)
pu.session_backup()
for sim in range(N_SIM):
# backup then make next move
for turn_sim in range(turn, N_TURNS+1):
for player in [0,1]:
# get valid moves, network policy and value estimates:
valid_mv_map, pol, val = arch.sess.run([arch.valid_mv_map, arch.pol, arch.val], feed_dict=ret_d(player))
# backup visit Q values
if turn_sim != turn:
pu.backup_visit(player, val)
pu.add_valid_mvs(player, valid_mv_map) # register valid moves in tree
to_coords = pu.choose_moves(player, pol, CPUCT)[0] # choose moves based on policy and Q values (latter of which already stored in tree)
pu.register_mv(player, to_coords) # register move in tree
arch.sess.run(arch.move_frm_inputs, feed_dict={arch.moving_player: player, arch.to_coords_input: to_coords}) # move network (update GPU vars)
# backup terminal state
for player in [0,1]:
winner = arch.sess.run(arch.winner, feed_dict=ret_d(player))
pu.backup_visit(player, winner)
# return move back to previous node in tree
arch.sess.run(arch.session_restore)
pu.session_restore()
def run_sim(turn, starting_player, scopes=['main', 'main']):
arch.sess.run(arch.session_backup)
pu.session_backup()
for sim in range(N_SIM):
# backup then make next move
for turn_sim in range(turn, N_TURNS + 1):
for player, s in zip([0, 1], scopes):
if turn_sim == turn and starting_player == 1 and player == 0: # skip player 0, has already moved
continue
# get valid moves, network policy and value estimates:
valid_mv_map, pol, val = arch.sess.run(
[arch.valid_mv_map, arch.pol[s], arch.val[s]],
feed_dict=ret_d(player))
# backup visit Q values
if turn_sim != turn:
pu.backup_visit(player, np.array(val, dtype='single'))
pu.add_valid_mvs(player,
valid_mv_map) # register valid moves in tree
to_coords = pu.choose_moves(
player, np.array(pol, dtype='single'), CPUCT
)[0] # choose moves based on policy and Q values (latter of which already stored in tree)
pu.register_mv(player, np.array(
to_coords, dtype='int32')) # register move in tree
arch.sess.run(arch.move_frm_inputs,
feed_dict={
arch.moving_player: player,
arch.to_coords_input: to_coords
}) # move network (update GPU vars)
############ backup terminal state
winner = np.array(arch.sess.run(arch.winner, feed_dict=ret_d(0)),
dtype='single')
# update tree with values (outcomes) of each game)
pu.backup_visit(0, winner)
pu.backup_visit(1, -winner)
# return move back to previous node in tree
arch.sess.run(arch.session_restore) # reset gpu game state
pu.session_restore() # reset cpu tree state
def nn_mv():
global Q_map, P_map, visit_count_map
global Q_map_next, P_map_next, visit_count_map_next
t_start = time.time()
arch.sess.run(arch.session_backup)
pu.init_tree()
pu.session_backup()
if run_net:
if turn == 0:
arch.sess.run(arch.nn_prob_move_unit_valid_mvs, feed_dict=ret_d(0))
else:
arch.sess.run(arch.nn_max_prob_move_unit_valid_mvs,
feed_dict=ret_d(0))
Q_map, P_map, visit_count_map = ret_stats(0)
else:
for sim in range(N_SIM):
# initial moves
for player in [0, 1]:
valid_mv_map, pol = arch.sess.run(
[arch.valid_mv_map, arch.pol], feed_dict=ret_d(player))
pu.add_valid_mvs(player, valid_mv_map)
to_coords = pu.choose_moves(player, pol, CPUCT)[0]
pu.register_mv(player, to_coords)
arch.sess.run(arch.move_frm_inputs,
feed_dict={
arch.moving_player: player,
arch.to_coords_input: to_coords
})
# backup then make next move
for turn_sim in range(turn, N_TURNS):
for player in [0, 1]:
valid_mv_map, pol, val = arch.sess.run(
[arch.valid_mv_map, arch.pol, arch.val],
feed_dict=ret_d(player))
pu.backup_visit(player, val)
pu.add_valid_mvs(player, valid_mv_map)
to_coords = pu.choose_moves(player, pol, CPUCT)[0]
pu.register_mv(player, to_coords)
arch.sess.run(arch.move_frm_inputs,
feed_dict={
arch.moving_player: player,
arch.to_coords_input: to_coords
})
# backup terminal state
for player in [0, 1]:
winner = arch.sess.run(arch.winner, feed_dict=ret_d(player))
pu.backup_visit(player, winner)
arch.sess.run(arch.session_restore)
pu.session_restore()
if sim % 20 == 0:
'''Q_map, P_map, visit_count_map = ret_stats(0)
arch.sess.run(arch.tree_det_move_unit, feed_dict = ret_d(0))
Q_map_next, P_map_next, visit_count_map_next = ret_stats(1)
arch.sess.run(arch.session_restore)
pu.session_restore()
draw(True)
pygame.display.set_caption('%i %2.1f' % (sim, time.time() - t_start))
'''
print 'simulation', sim, 'total elapsed time', time.time(
) - t_start
### make move
Q_map, P_map, visit_count_map = ret_stats(0)
valid_mv_map, pol = arch.sess.run(
[arch.imgs, arch.valid_mv_map, arch.pol], feed_dict=ret_d(0))[1:]
#########
pu.add_valid_mvs(player, valid_mv_map)
visit_count_map = pu.choose_moves(player, pol, CPUCT)[-1]
to_coords = arch.sess.run(
[arch.tree_det_visit_coord, arch.tree_det_move_unit],
feed_dict={
arch.moving_player: 0,
arch.visit_count_map: visit_count_map,
arch.dir_pre: dir_pre,
arch.dir_a: DIR_A
})[0]
pu.register_mv(player, to_coords)
pu.prune_tree()
print time.time() - t_start
return arch.sess.run(arch.gm_vars['board'])[0]
def nn_mv():
global Q_map, P_map, visit_count_map, valid_mv_map, pol
global Q_map_next, P_map_next, visit_count_map_next, to_coords
t_start = time.time()
arch.sess.run(arch.session_backup)
#### make most probable mv, do not use tree search
if run_one_pass_only:
# 'eval32' movement ops were not defined, so get policy, from network, and then use the ops in 'eval' (where it was defined)
d = ret_d(NET_PLAYER)
imgs = arch.sess.run(arch.imgs, feed_dict=d)
d[arch.imgs32] = np.asarray(imgs, dtype='float')
pol = arch.sess.run(arch.pol[net], feed_dict=d)
d = ret_d(NET_PLAYER)
d[arch.pol['eval']] = pol
if turn == 0:
arch.sess.run(arch.nn_prob_move_unit_valid_mvs['eval'],
feed_dict=d)
else:
arch.sess.run(arch.nn_max_prob_move_unit_valid_mvs['eval'],
feed_dict=d)
#Q_map, P_map, visit_count_map = ret_stats(0)
##### use tree search
else:
#pu.init_tree()
pu.session_backup()
sim = 0
# each loop is one simulation
while True:
if ((time.time() - t_start) > TIME_MIN) and (sim >= SIM_MIN):
break
# backup then make next move
# (this loop, iterates over one full game-play from present turn)
for turn_sim in range(turn, np.max(
(N_TURNS + 1, turn + TURN_MIN))):
for player in [0, 1]:
if turn_sim == turn and human_player(
) == 0 and player == 0: # skip player 0 (human), has already moved
continue
# get valid moves, network policy and value estimates:
valid_mv_map, pol, val = arch.sess.run(
[arch.valid_mv_map, arch.pol[net], arch.val[net]],
feed_dict=ret_d(player))
# backup visit Q values
if turn_sim != turn:
pu.backup_visit(player, np.array(val, dtype='single'))
pu.add_valid_mvs(
player, valid_mv_map) # register valid moves in tree
to_coords = pu.choose_moves(
player, np.array(pol, dtype='float32'), CPUCT
)[0] # choose moves based on policy and Q values (latter of which already stored in tree)
pu.register_mv(player, np.array(
to_coords, dtype='int32')) # register move in tree
arch.sess.run(arch.move_frm_inputs,
feed_dict={
arch.moving_player: player,
arch.to_coords_input: to_coords
}) # move network (update GPU vars)
# backup terminal state
winner = np.array(arch.sess.run(arch.winner, feed_dict=ret_d(0)),
dtype='single')
pu.backup_visit(0, winner)
pu.backup_visit(1, -winner)
# return move to previous node in tree
arch.sess.run(arch.session_restore) # reset gpu game state
pu.session_restore() # reset cpu tree state
######################
# print stats from tree
if sim % 20 == 0:
# get valid moves, network policy and value estimates:
valid_mv_map = arch.sess.run([arch.imgs, arch.valid_mv_map],
feed_dict=ret_d(NET_PLAYER))[1]
pu.add_valid_mvs(NET_PLAYER,
valid_mv_map) # register valid moves in tree
visit_count_map_128 = pu.choose_moves(
NET_PLAYER, np.array(pol, dtype='float32'), CPUCT
)[-1] # to feed back into tf (entries for all 128 games, not just 1)
Q_map, P_map, visit_count_map = ret_stats(
NET_PLAYER) # stats we will show on screen
# move network where it is estimates is its best move
to_coords = arch.sess.run(
[
arch.nn_max_prob_to_coords_valid_mvs[net],
arch.nn_max_prob_move_unit_valid_mvs[net]
],
feed_dict={
arch.moving_player: NET_PLAYER,
arch.pol[net]: visit_count_map_128
})[0]
pu.register_mv(NET_PLAYER, np.asarray(
to_coords, dtype='int32')) # register move in tree
arch.sess.run(arch.move_frm_inputs,
feed_dict={
arch.moving_player: NET_PLAYER,
arch.to_coords_input: to_coords
}) # move network (update GPU vars)
# get network tree estimates as to where it thinks you will move after it moves
valid_mv_map = arch.sess.run([arch.imgs, arch.valid_mv_map],
feed_dict=ret_d(
human_player()))[1]
pu.add_valid_mvs(human_player(),
valid_mv_map) # register valid moves in tree
Q_map_next, P_map_next, visit_count_map_next = ret_stats(
human_player())
arch.sess.run(
arch.session_restore) # restore prior tf game state
pu.session_restore() # restore prior tree
draw(True)
pygame.display.set_caption('%i %2.1f' %
(sim, time.time() - t_start))
print 'simulation: ', sim, ' (%i sec)' % (time.time() -
t_start)
sim += 1
### make move
# first get valid moves and current policy at board position
valid_mv_map, pol = arch.sess.run(
[arch.imgs, arch.valid_mv_map, arch.pol[net]],
feed_dict=ret_d(NET_PLAYER))[1:]
pu.add_valid_mvs(NET_PLAYER, valid_mv_map) # set in tree
visit_count_map_128 = pu.choose_moves(
NET_PLAYER, np.array(pol, dtype='float32'), CPUCT
)[-1] # to feed back into tf (entries for all 128 games, not just 1)
Q_map, P_map, visit_count_map = ret_stats(NET_PLAYER)
# makes moves as if this were still part of the self-play (max visit count)
#to_coords = arch.sess.run([arch.tree_det_visit_coord, arch.tree_det_move_unit], feed_dict={arch.moving_player: 0,
# arch.visit_count_map: visit_count_map})[0]
# move to max visited node:
#if turn != 0:
to_coords = arch.sess.run([
arch.nn_max_prob_to_coords_valid_mvs[net],
arch.nn_max_prob_move_unit_valid_mvs[net]
],
feed_dict={
arch.moving_player: NET_PLAYER,
arch.pol[net]: visit_count_map_128
})[0]
# randomly move proportionatly to vist counts
#else:
# to_coords = arch.sess.run([arch.tree_prob_visit_coord, arch.tree_prob_move_unit], feed_dict={arch.moving_player: 0,
# arch.visit_count_map: visit_count_map})[0] # make move in proportion to visit counts
pu.register_mv(NET_PLAYER, np.array(to_coords, dtype='int32'))
print 'pruning...'
pu.prune_tree(
1) # 0: prune all games in batch, 1: prune only first game
print time.time() - t_start
print 'finished'
return arch.sess.run(arch.gm_vars['board'])[0]