def evaluate_candidate(candidate_player, reference_player):
wins = 0
draws = 0
losses = 0
# Always choose best move
candidate_player.tryhard_mode = True
for _ in range(400):
env = GameEnv()
current_player = candidate_player if random.random() < 0.5 else reference_player
while True:
move = current_player.select_move(env)
env.play_move(*move) # Current player flips
env.check_win()
if env.winner and current_player is candidate_player:
wins += 1
break
elif env.winner and current_player is reference_player:
losses += 1
break
elif env.draw:
draws += 1
break
else:
current_player = reference_player if current_player is candidate_player else candidate_player
candidate_player.tryhard_mode = False
return wins, draws, losses
def main():
model_name = "models/model_conv.h5"
reference_player = TinyBrain()
# candidate_player = BigBrain(load_model=model_name, tryhard_mode=False) # For continuing training
candidate_player = BigBrain(tryhard_mode=False) # For starting training
episode = 1
while True:
env = GameEnv()
first_move = True
# immediate store is all the information we have immediately after a move
# (current state, possible actions, move)
immediate_store = []
# delayed store is all the information we get after the next move
# (next state, reward, terminated)
delayed_store = []
# Randomly choose who goes first
current_player = candidate_player if random.random() < 0.5 else reference_player
while True:
state = env.state()
move = current_player.select_move(env)
if current_player is candidate_player:
if first_move:
first_move = False
else:
# Finish providing information for candidate player's last move
possible_actions = env.possible_actions()
delayed_store.append((possible_actions, 0, state, False))
# Provide starting information for candidate player's current move
do_explore = random.random() < 0.3
move = current_player.select_move(env, explore=do_explore)
immediate_store.append((state, move, do_explore))
env.play_move(*move) # Current player flips
env.check_win()
if env.winner or env.draw:
# If game has ended we need to give rewards to both players
if env.draw:
delayed_store.append((None, DRAW_REWARD, None, True))
elif current_player is candidate_player:
# Winner is always whoever played last
delayed_store.append((None, WIN_REWARD, None, True))
else:
delayed_store.append((None, LOSS_REWARD, None, True))
for immediate, delayed in zip(immediate_store, delayed_store):
state, move, do_explore = immediate
if not do_explore:
candidate_player.store(state, move, *delayed)
break
current_player = reference_player if current_player is candidate_player else candidate_player
if episode % GAMES_PER_SET == 0:
print(f"Training after {episode} episodes")
candidate_player.retrain(batch_size=TRAINING_SIZE)
candidate_player.align_target_model()
candidate_player.save(model_name)
wins, draws, losses = evaluate_candidate(candidate_player, TinyBrain())
print(f"{wins}, {draws}, {losses}")
if wins + losses > 0:
percentage_wins = wins / (wins + losses)
else:
percentage_wins = 0
print(f"percentage wins: {percentage_wins}")
reference_player = candidate_player
reference_player.tryhard_mode = False
candidate_player = BigBrain(tryhard_mode=False)
candidate_player.q_network = reference_player.q_network
candidate_player.align_target_model()
episode += 1