def multisine_generate(system,
task,
dynamics,
rng,
init_min,
init_max,
n_freqs,
traj_len,
n_trajs,
abort_if=None):
trajs = []
periods = list(range(1, traj_len, n_freqs))
umin, umax = task.get_ctrl_bounds().T
uamp = (umax - umin) / 2
umed = (umax + umin) / 2
for _ in range(n_trajs):
weights = []
for i in range(system.ctrl_dim):
vals = rng.uniform(size=len(periods) - 1)
vals = np.concatenate([[0.0], np.sort(vals), [1.0]])
weight = vals[1:] - vals[:-1]
weights.append(weight)
weights = np.array(weights)
phases = rng.uniform(0, 2 * np.pi, len(periods))
state0 = [
rng.uniform(minval, maxval, 1)[0]
for minval, maxval in zip(init_min, init_max)
]
y = state0[:]
traj = ampc.zeros(system, traj_len)
traj.obs[:] = y
umin, umax = task.get_ctrl_bounds().T
for i in range(traj_len):
traj[i].obs[:] = y
u = np.zeros(system.ctrl_dim)
for j, period in enumerate(periods):
u += weights[:, j] * np.cos(2 * np.pi * i / period + phases[j])
u = uamp * u + umed
y = dynamics(y, u)
traj[i].ctrl[:] = u
if not abort_if is None and abort_if(y):
traj = traj[:i]
break
trajs.append(traj)
return trajs
def uniform_random_generate(system, task, dynamics, rng, init_min, init_max,
traj_len, n_trajs):
trajs = []
for _ in range(n_trajs):
state0 = [
rng.uniform(minval, maxval, 1)[0]
for minval, maxval in zip(init_min, init_max)
]
y = state0[:]
traj = ampc.zeros(system, traj_len)
traj.obs[:] = y
umin, umax = task.get_ctrl_bounds().T
for i in range(traj_len):
traj[i].obs[:] = y
u = rng.uniform(umin, umax, system.ctrl_dim)
y = dynamics(y, u)
traj[i].ctrl[:] = u
trajs.append(traj)
return trajs
def periodic_control_generate(system, task, dynamics, rng, init_min, init_max,
U_1, traj_len, n_trajs):
trajs = []
periods = list(range(1, traj_len, max([1, traj_len // n_trajs])))
print("periods=", periods)
for period in periods:
state0 = [
rng.uniform(minval, maxval, 1)[0]
for minval, maxval in zip(init_min, init_max)
]
y = state0[:]
traj = ampc.zeros(system, traj_len)
traj.obs[:] = y
umin, umax = task.get_ctrl_bounds().T
uamp = np.min([umin, umax])
for i in range(traj_len):
traj[i].obs[:] = y
u = uamp * U_1 * np.cos(2 * np.pi * i / period)
y = dynamics(y, u)
traj[i].ctrl[:] = u
trajs.append(traj)
return trajs
def random_walk_generate(system, task, dynamics, rng, init_min, init_max,
walk_rate, traj_len, n_trajs):
trajs = []
for _ in range(n_trajs):
state0 = [
rng.uniform(minval, maxval, 1)[0]
for minval, maxval in zip(init_min, init_max)
]
y = state0[:]
traj = ampc.zeros(system, traj_len)
traj.obs[:] = y
umin, umax = task.get_ctrl_bounds().T
uamp = np.min([umin, umax])
u = rng.uniform(umin, umax, system.ctrl_dim)
step_size = walk_rate * system.dt
for i in range(traj_len):
traj[i].obs[:] = y
u += uamp * step_size * rng.uniform(-1, 1, system.ctrl_dim)
u = np.clip(u, umin, umax)
y = dynamics(y, u)
traj[i].ctrl[:] = u
trajs.append(traj)
return trajs
def prbs_generate(system, task, dynamics, rng, init_min, init_max, traj_len,
n_trajs, states, Nswitch):
trajs = []
for _ in range(n_trajs):
# Compute control sequence
switches = rng.choice(traj_len, Nswitch)
switches = np.concatenate([[0], switches, [traj_len]])
u = np.zeros(traj_len)
for ps, ns in zip(switches[:-1], switches[1:]):
u[ps:ns] = rng.choice(states)
state0 = [
rng.uniform(minval, maxval, 1)[0]
for minval, maxval in zip(init_min, init_max)
]
y = state0[:]
traj = ampc.zeros(system, traj_len)
traj.obs[:] = y
for i in range(traj_len):
traj[i].obs[:] = y
y = dynamics(y, u[i])
traj[i].ctrl[:] = u[i]
trajs.append(traj)
return trajs