def self_play(storage, player1, player2=None, explore=True, num_games=1, joseki=False):
for n in range(num_games):
if not parallell:
print("Self-play game: %s" % n)
#Handle the fact that ordinary self-play uses a single tree structure
#whereas evaluation uses two different ones
if player2 != None: evaluation = True
else: evaluation = False
#Initialize game structure
game = santorini.Game()
#Initialize players with networks and tree structures. Make the structures
#globally available to facilitiate inspection or debugging
p1 = M.MCTS(game, player1, sess, explore)
global P1
P1 = p1
if player2 != None:
evaluation = True
p2 = M.MCTS(game, player2, sess, explore)
global P2
P2 = p2
players = [p1, p2]
else:
evaluation = False
#Store state history, but don't add it to global history yet as we need
#to know the outcome first
temp_history = []
done = False
while done == False:
if evaluation:
player = game.turn_count%2
tree = players[player]
other_tree = players[(player+1)%2]
else:
tree = p1
#Execute tree search and make move
t0 = time.time()
done = tree.consider_resigning(v_resign, observe_games)
a, pi_s, P, v = tree.run_simulation(search_depth)
temp_history.extend([[game.stack_s(), pi_s, game.legal_moves(binaryV=True)]])
if evaluation: #This is not very neat, and I should fix it up at some point...
other_tree.prepare_adversarial_move(a)
game.move(a)
done = game.done
if evaluation:
other_tree.finish_adversarial_move(a)
tree.prepare_next_move()
if observe_games:
for i in range(10):
print("\n")
print("P (predicted tree search probs):\n%s\n\n" % np.reshape(P, [5,5]),
"pi (actual tree search probs):\n%s\n\n" % np.reshape(pi_s, [5,5]),
"v: %s\n" % v,
"Chosen move: %s\n" % a,
"Overall game state:\n%s\n\n" % game.render())
print("time: ", time.time()-t0)
z = game.outcome
#store data
t = len(temp_history)
for entry in temp_history:
storage.add(entry[0], entry[1], discount_rs(z, t), entry[2])
t -= 1
return z