nb_states = 2
nb_actions = 2
WB_model = WB_Model(dt)
BB_model = BB_Model(nb_states, nb_actions)
loaded = BB_model.load_model(path='linear_2D_')
if loaded: print("2 Input Linear model loaded!")
# create the SUSD learner
Q = np.eye(2)
R = np.eye(2)
K0 = np.array([[-10., -10.],[-10., -10.]])
N_trajectories = 10
N_agents = 25
T_rollout = 10
S = SUSD(WB_model, Q, R, K0.copy(), N_agents, T_rollout)
PG = Policy_Gradient(WB_model, Q, R, K0.copy(), N_trajectories, T_rollout)
# compute the optimal K
A, B = WB_model.system()
K_lqr, _, _ = control.lqr(A, B, Q, R)
# get optimal K cost
C_lqr = 0
x = x0.copy()
Nsteps = int(T_rollout/dt)
for t in range(Nsteps):
# compute input
u = np.dot(-K_lqr, x)
# compute cost
for idn in range(len(n)):
# define simulation length
Nsteps = 500
# define starting state
x0 = np.array([-8, 9])
# load model predictor
nb_states = 2
nb_actions = 1
wb_model = WB_Model(dt)
# create the SUSD learner
Q = np.eye(2)
R = np.array([1])
K0 = np.array([0.1, 0.1])
N_rollout = n[idn]
T_rollout = 20
S = SUSD(wb_model, Q, R, K0, N_rollout, T_rollout, alpha=0.35, term_len=100)
# estimate optimal K using SUSD
conveged, iters = S.search(x0, max_iter=100)
K = S.K.reshape((nb_actions, nb_states))
print "SUSD Search took", iters, "iterations"
print "Estimated K:"; print K
z_list[idn] = S.z_buf
cost_plot(z_list, n)
plt.show()
nb_states = 4
nb_actions = 2
WB_model = WB_Model(dt)
# create the SUSD learner
Q = np.eye(4)
R = np.eye(2)
K0 = np.array([[-2., -2., -2., -2.], [-2., -2., -2., -2.]])
N_agents = 30
T_rollout = 10
S = SUSD(WB_model,
Q,
R,
K0,
N_agents,
T_rollout,
dt=dt,
alpha=0.1,
term_len=200)
# compute the optimal K
A, B = WB_model.system()
sys = control.StateSpace(A, B, np.eye(4), np.zeros((4, 2)), 1)
K_lqr, _, _ = control.lqr(sys, Q, R)
# estiamte optimal K using SUSD
conveged, iters = S.search(x0, r=0.001, max_iter=100000)
K = S.K.reshape((nb_actions, nb_states))
print "SUSD Search took", iters, "iterations"
R = np.array([1])
K0 = np.array([-10., -10.])
N_points = 10
N_trajectories = 100
N_agents = 5
T_rollout = 10
RS = Random_Search(WB_model,
Q,
R,
K0.copy(),
N_points,
T_rollout,
0.01 * np.eye(2),
dt=dt)
S = SUSD(WB_model, Q, R, K0.copy(), N_agents, T_rollout, dt=dt, alpha=0.9)
PG = Policy_Gradient(WB_model,
Q,
R,
K0.copy(),
N_trajectories,
T_rollout,
dt=dt,
alpha=9E-1)
# compute the optimal K
A, B = WB_model.system()
K_lqr, _, _ = control.lqr(A, B, Q, R)
print "LQR Gains:", -K_lqr
# estimate optimal K using FD
N_agents = 30
T_rollout = 10
RS = Random_Search(WB_model,
Q,
R,
K0.copy(),
N_points,
T_rollout,
0.5 * np.eye(12),
dt=dt)
S = SUSD(WB_model,
Q,
R,
K0.copy(),
N_agents,
T_rollout,
dt=dt,
alpha=0.1,
term_len=100)
PG = Policy_Gradient(WB_model,
Q,
R,
K0.copy(),
N_trajectories,
T_rollout,
dt=dt,
alpha=0.8)
# compute the optimal K
A, B = WB_model.system()
# load model predictor nb_states = 2 nb_actions = 1 model = BB_Model(nb_states, nb_actions) model.load_model(path='linear_said_') WB_model = WB_Model(dt) # create the SUSD learner Q = np.eye(2) R = np.array([1]) K0 = np.array([-10., -10.]) N_rollout = 3 T_rollout = 30 S = SUSD(WB_model, Q, R, K0, N_rollout, T_rollout) # compute the optimal K A, B = WB_model.system() K_lqr, _, _ = control.lqr(A, B, Q, R) # estiamte optimal K using SUSD conveged, iters = S.search(x0) K = S.K.reshape((nb_actions, nb_states)) print "SUSD Search took", iters, "iterations" print "Estimated K:", K print "LQR K:", -K_lqr # simulate the trajectory x = np.zeros([nb_states, Nsteps])
# load model predictor
nb_states = 2
nb_actions = 2
bb_model = BB_Model(nb_states, nb_actions)
loaded = bb_model.load_model(path='linear_2D_')
if loaded: print("2 Input Linear model loaded!")
WB_model = WB_Model(dt)
# create the SUSD learner
Q = np.eye(2)
R = np.eye(2)
K0 = np.array([[-1., -1.], [-1., -1.]])
N_agents = 10
T_rollout = 15
S = SUSD(bb_model, Q, R, K0, N_agents, T_rollout)
# compute the optimal K
A, B = WB_model.system()
K_lqr, _, _ = control.lqr(A, B, Q, R)
# estiamte optimal K using SUSD
conveged, iters = S.search(x0)
K = S.K.reshape((nb_actions, nb_states))
print "SUSD Search took", iters, "iterations"
print "Estimated K:", K
print "LQR K:", -K_lqr
# simulate the trajectory
x = np.zeros([nb_states, Nsteps])