def do_rollouts(self, seed):
paths = gather_paths_parallel(self.env.env_id,
self.sol_state[-1],
self.act_sequence,
self.filter_coefs,
seed,
self.paths_per_cpu,
self.num_cpu,
)
return paths
def do_rollouts(self, seed, goal=None):
paths = gather_paths_parallel(self.env.env_name,
self.sol_state[-1],
self.act_sequence,
self.filter_coefs,
seed,
goal,
self.reward_type,
self.reference,
self.env.alpha,
self.paths_per_cpu,
self.num_cpu,
)
return paths
def get_action(self, env_state):
# Set to env_state
# Shoot trajectories
# Return optimal action
seed = self.seed + self.ctr * 1000
paths = gather_paths_parallel(self.env.env_id,
env_state,
self.act_sequence,
self.filter_coefs,
seed,
self.paths_per_cpu,
self.num_cpu,
)
num_traj = len(paths)
R = self.score_trajectory(paths)
S = np.exp(self.kappa*(R-np.max(R)))
act = np.sum([paths[i]["actions"][0] * S[i] for i in range(num_traj)], axis=0)
act = act / (np.sum(S) + 1e-6)
return act