def __init__(self):
self.env = cpommerman.make()
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {self.device}")
self.net = Net().to(self.device)
self.net.eval()
self.optimizer = torch.optim.Adam(self.net.parameters(), lr=3e-4)
self.oracle = FDTS(100, 20)
self.buffer = deque(maxlen=10**5)
def runner(id, num_episodes, args, fifo):
env = cpommerman.make()
if args.planner == 'fdts':
planner = FDTS(args.n_simulations, args.horizon, args.mab)
elif args.planner == 'mcts':
planner = MCTS(args.n_simulations, args.mab)
elif args.planner == 'mcs':
planner = MCS(args.n_simulations, args.horizon, args.mab)
else:
raise Exception('Unknown planner')
outcomes = [0, 0, 0] # wins, draws, losses
for i in range(num_episodes):
env.reset()
planner.reset()
planner_id = np.random.randint(4)
simple = JointSimpleAgent(planner_id)
while not env.get_done():
obses = env.get_observations()
actions = simple.step(obses)
planner_actions = planner.step(env, env.get_legal_actions())
actions[planner_id] = planner_actions[planner_id]
state_str = planner.get_state_str(env.get_json_info())
env.step(actions)
rewards = env.get_rewards()
if rewards[planner_id] == 1: # win
idx = 0
elif sum(rewards) == -4: # draw
idx = 1
else: # loss
idx = 2
outcomes[idx] += 1
fifo.put(idx)
parser = argparse.ArgumentParser()
parser.add_argument('--num_episodes', type=int, default=1)
parser.add_argument('--render', action="store_true", default=False)
args = parser.parse_args()
# set up two environments, agents are only in old env
agents = [
agents.SimpleAgent(),
agents.SimpleAgent(),
agents.SimpleAgent(),
agents.SimpleAgent(),
]
env_old = pommerman.make('PommeFFACompetition-v0', agents)
env_new = cpommerman.make()
total_time_old = 0
total_time_new = 0
total_state_len = 0
max_state_len = 0
n = 0
for i in range(args.num_episodes):
obs_old = env_old.reset()
# just to test for errors in reset code,
# impossible to verify because random seed is not implemented in both envs
obs_new = env_new.reset()
done_old = False
t = 0
while not done_old:
def __init__(self, model_file, agent_id=0):
super().__init__()
self.model = load_model(model_file)
self.agent_id = agent_id
self.env = cpommerman.make()
self.reset_tree()
def __init__(self, model, agent_id=0):
super().__init__()
self.model = model
self.agent_id = agent_id
self.env = cpommerman.make()
self.reset_tree()
import cpommerman
import numpy as np
import time
env = cpommerman.make()
start_time = time.time()
steps = 0
for i in range(1000):
env.reset()
done = False
while not done:
#state = env.get_state()
#obs = env.get_observations()
features = env.get_features()
# use features, observations or state to produce action
actions = np.random.randint(6, size=4, dtype=np.uint8)
env.step(actions)
rewards = env.get_rewards()
done = env.get_done()
steps += 1
elapsed = time.time() - start_time
print("Time:", elapsed, "Steps:", steps, "Time per step:", elapsed / steps, "FPS:", steps / elapsed)