inner_env = normalize(
AntMazeEnv(
maze_id=par.maze_id,
death_reward=par.death_reward,
sensor_range=par.sensor_range,
sensor_span=math.pi * 2,
ego_obs=True,
fence=par.fence,
goal_rew=par.success_reward,
random_start=par.random_start,
direct_goal=par.direct_goal,
velocity_field=par.velocity_field,
))
env = hierarchize_snn(
inner_env,
time_steps_agg=time_step_agg,
pkl_path=par.pkl_path,
animate=par.animate,
)
policy = CategoricalMLPPolicy(env_spec=env.spec, )
print("env_hier", env.spec)
if par.baseline_name == 'linear':
baseline = LinearFeatureBaseline(env_spec=env.spec)
elif par.baseline_name == 'mlp':
baseline = GaussianMLPBaseline(env_spec=env.spec)
batch_size = int(par.low_step_num / time_step_agg)
max_path_length = int(par.max_low_step / time_step_agg)
algo = TRPO(
env=env,
mode = "local"
n_parallel = 4
exp_dir = 'data/local/egoSwimmer-snn/'
for dir in os.listdir(exp_dir):
if 'Figure' not in dir and os.path.isfile(os.path.join(exp_dir, dir, 'params.pkl')):
pkl_path = os.path.join(exp_dir, dir, 'params.pkl')
print("hier for : ", pkl_path)
for time_step_agg in [10, 50, 100]:
for activity_range in [6, 10, 15]:
inner_env = normalize(SwimmerGatherEnv(activity_range=activity_range, sensor_range=activity_range,
sensor_span=math.pi * 2, ego_obs=True))
env = hierarchize_snn(inner_env, time_steps_agg=time_step_agg, pkl_path=pkl_path,
# animate=True,
)
policy = CategoricalMLPPolicy(
env_spec=env.spec,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
# bonus_evaluators = [GridBonusEvaluator(mesh_density=mesh_density, visitation_bonus=1, snn_H_bonus=0)]
# reward_coef_bonus = [reward_coef]
algo = TRPO_snn(
env=env,
policy=policy,
baseline=baseline,