# setup the environment
env = MultiObjectiveEnv(args.env_name)
# generate an agent for plotting
agent = None
if args.method == 'crl-naive':
from crl.naive.meta import MetaAgent
elif args.method == 'crl-envelope':
from crl.envelope.meta import MetaAgent
elif args.method == 'crl-energy':
from crl.energy.meta import MetaAgent
model = torch.load("{}{}.pkl".format(
args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name,
args.name)))
agent = MetaAgent(model, args, is_train=False)
# compute opt
opt_x = []
opt_y = []
q_x = []
q_y = []
act_x = []
act_y = []
real_sol = np.stack((dis_treasure, dis_time))
policy_loss = np.inf
predict_loss = np.inf
for i in range(2000):
w = np.random.randn(2)
# args.episode_num = 600
# setup the environment
env = MultiObjectiveEnv(args.env_name)
# get state / action / reward sizes
state_size = len(env.state_spec)
action_size = env.action_spec[2][1] - env.action_spec[2][0]
reward_size = len(env.reward_spec)
# generate an agent for initial training
agent = None
if args.method == 'crl-naive':
from crl.naive.meta import MetaAgent
from crl.naive.models import get_new_model
elif args.method == 'crl-envelope':
from crl.envelope.meta import MetaAgent
from crl.envelope.models import get_new_model
elif args.method == 'crl-energy':
from crl.energy.meta import MetaAgent
from crl.energy.models import get_new_model
if args.serialize:
model = torch.load("{}{}.pkl".format(
args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name,
args.name)))
else:
model = get_new_model(args.model, state_size, action_size, reward_size)
agent = MetaAgent(model, args, is_train=True)
train(env, agent, args)
# setup the environment
env = MultiObjectiveEnv(args.env_name)
# generate an agent for plotting
agent = None
if args.method == 'crl-naive':
from crl.naive.meta import MetaAgent
elif args.method == 'crl-envelope':
from crl.envelope.meta import MetaAgent
elif args.method == 'crl-energy':
from crl.energy.meta import MetaAgent
model = torch.load("{}{}.pkl".format(
args.save, "m.{}_e.{}_n.{}".format(args.model, args.env_name, args.name)))
agent = MetaAgent(model, args, is_train=False)
REPEAT = 1
# unknown pref
unknown_w = np.array([0.0, 0.0, 0.0, 0.0, 1.0, 0.0])
bbbest_param = []
for iii in range(5):
pref_param = np.array(
[1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
explore_w = generate_w(1, pref_param)
max_target = 0
sample_episode = 0