def setup_opponents(self):
if self.opponent_type == 'rules':
self.opponent_agent = Agent('greedy')
else:
# incremental load of new model
best_model_name = get_best_model_name(self.name)
if self.best_model_name != best_model_name:
self.opponent_models.append(
load_model(self, best_model_name))
self.best_model_name = best_model_name
if self.opponent_type == 'random':
start = 0
end = len(self.opponent_models) - 1
i = random.randint(start, end)
self.opponent_agent = Agent('ppo_opponent',
self.opponent_models[i])
elif self.opponent_type == 'best':
self.opponent_agent = Agent('ppo_opponent',
self.opponent_models[-1])
elif self.opponent_type == 'mostly_best':
j = random.uniform(0, 1)
if j < 0.8:
self.opponent_agent = Agent('ppo_opponent',
self.opponent_models[-1])
else:
start = 0
end = len(self.opponent_models) - 1
i = random.randint(start, end)
self.opponent_agent = Agent('ppo_opponent',
self.opponent_models[i])
elif self.opponent_type == 'base':
self.opponent_agent = Agent('base',
self.opponent_models[0])
self.agent_player_num = np.random.choice(self.n_players)
self.agents = [self.opponent_agent] * self.n_players
self.agents[self.agent_player_num] = None
try:
# if self.players is defined on the base environment
logger.debug(
f'Agent plays as Player {self.players[self.agent_player_num].id}'
)
except:
pass
def __init__(self, env, name="AI_PPO", size=6):
super().__init__(name)
# make environment
ppo_model = load_model(env, f'best_model{size}.zip')
self.ppo_agent = Agent('best_model', ppo_model)
def main(args):
logger.configure(config.LOGDIR)
if args.debug:
logger.set_level(config.DEBUG)
else:
logger.set_level(config.INFO)
# make environment
env = get_environment(args.env_name)(verbose=args.verbose,
manual=args.manual)
env.seed(args.seed)
total_rewards = {}
first_time = True
if args.recommend:
ppo_model = load_model(env, 'best_model.zip')
ppo_agent = Agent('best_model', ppo_model)
else:
ppo_agent = None
agents = []
# load the agents
if len(args.agents) != env.n_players:
raise Exception(
f'{len(args.agents)} players specified but this is a {env.n_players} player game!'
)
for i, agent in enumerate(args.agents):
if agent == 'human':
agent_obj = Agent('human')
elif agent == 'rules':
agent_obj = Agent('rules')
elif agent == 'json':
# Start mq server
context = zmq.Context()
socket = context.socket(zmq.REP)
socket.bind("tcp://*:5555")
logger.debug("zaq server start at 5555")
agent_obj = Agent('json')
elif agent == 'base':
base_model = load_model(env, 'base.zip')
agent_obj = Agent('base', base_model)
else:
ppo_model = load_model(env, f'{agent}.zip')
agent_obj = Agent(agent, ppo_model)
agents.append(agent_obj)
total_rewards[agent_obj.id] = 0
# play games
logger.info(f'\nPlaying {args.games} games...')
for game in range(args.games):
players = agents[:]
if args.randomise_players:
random.shuffle(players)
obs = env.reset()
done = False
for i, p in enumerate(players):
logger.debug(f'Player {i+1} = {p.name}')
while not done:
current_player = players[env.current_player_num]
env.render()
logger.debug(f'\nCurrent player name: {current_player.name}')
if args.recommend and current_player.name in [
'human', 'rules', 'json'
]:
# show recommendation from last loaded model
logger.debug(f'\nRecommendation by {ppo_agent.name}:')
action = ppo_agent.choose_action(env,
choose_best_action=True,
mask_invalid_actions=True)
if current_player.name == 'human':
action = input('\nPlease choose an action: ')
try:
# for int actions
action = int(action)
except:
# for MulitDiscrete action input as list TODO
action = eval(action)
if current_player.name == 'json':
if (not first_time):
game_state = {
"legal_action":
[i for i, o in enumerate(env.legal_actions) if o != 0],
"tableCard":
env.tableCard.id
}
socket.send_json(game_state)
action = socket.recv_json()
first_time = False
logger.debug(f'\nReceived {action}')
# action = input('\n JSON!!! Please choose an action: ')
try:
# for int actions
action = int(action)
except:
# for MulitDiscrete action input as list TODO
action = eval(action)
elif current_player.name == 'rules':
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(
env, choose_best_action=False, mask_invalid_actions=True)
else:
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(
env,
choose_best_action=args.best,
mask_invalid_actions=True)
obs, reward, done, _ = env.step(action)
for r, player in zip(reward, players):
total_rewards[player.id] += r
player.points += r
if args.cont:
input('Press any key to continue')
env.render()
logger.info(f"Played {game + 1} games: {total_rewards}")
if args.write_results:
write_results(players, game, args.games, env.turns_taken)
for p in players:
p.points = 0
env.close()
def main(args):
logger.configure(config.LOGDIR)
if args.debug:
logger.set_level(config.DEBUG)
else:
logger.set_level(config.INFO)
#make environment
env = get_environment(args.env_name)(verbose = args.verbose, manual = args.manual)
env.seed(args.seed)
total_rewards = {}
if args.recommend:
ppo_model = load_model(env, 'best_model.zip')
ppo_agent = Agent('best_model', ppo_model)
else:
ppo_agent = None
agents = []
#load the agents
if len(args.agents) != env.n_players:
raise Exception(f'{len(args.agents)} players specified but this is a {env.n_players} player game!')
for i, agent in enumerate(args.agents):
if agent == 'human':
agent_obj = Agent('human')
elif agent== 'greedy':
agent_obj = Agent('greedy')
elif agent == 'rules':
agent_obj = Agent('rules')
elif agent == 'base':
base_model = load_model(env, 'base.zip')
agent_obj = Agent('base', base_model)
else:
ppo_model = load_model(env, f'{agent}.zip')
agent_obj = Agent(agent, ppo_model)
agents.append(agent_obj)
total_rewards[agent_obj.id] = 0
if args.env_name == "blobwar":
human_blobwar = Human()
#play games
logger.info(f'\nPlaying {args.games} games...')
for game in range(args.games):
players = agents[:]
if args.randomise_players:
random.shuffle(players)
obs = env.reset()
done = False
for i, p in enumerate(players):
logger.debug(f'Player {i+1} = {p.name}')
while not done:
current_player = players[env.current_player_num]
env.render()
logger.debug(f'Current player name: {current_player.name}')
if args.recommend and current_player.name in ['human', 'rules']:
# show recommendation from last loaded model
logger.debug(f'\nRecommendation by {ppo_agent.name}:')
action = ppo_agent.choose_action(env, choose_best_action = True, mask_invalid_actions = True)
if current_player.name == 'human':
if args.env_name == "blobwar":
move= human_blobwar.compute_next_move(env.core)
action=env.encode_action(move)
else:
action = input('\nPlease choose an action: ')
try:
action = int(action)
except:
# for MulitDiscrete action input as list TODO
action = eval(action)
elif current_player.name == 'rules':
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(env, choose_best_action = False, mask_invalid_actions = True)
else:
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(env, choose_best_action = args.best, mask_invalid_actions = True)
obs, reward, done, _ = env.step(action)
for r, player in zip(reward, players):
total_rewards[player.id] += r
player.points += r
if args.cont:
input('Press any key to continue')
env.render()
logger.info(f"Played {game + 1} games: {total_rewards}")
if args.write_results:
write_results(players, game, args.games, env.turns_taken)
for p in players:
p.points = 0
env.close()