def worldA():
T = getT()
state_features = np.zeros((25, 3))
for x in xrange(5):
for y in xrange(5):
state_features[5*x + y] = np.array([0,1,0])
for x in xrange(5):
for y in xrange(5):
if x == 0 or y == 0 or x == 4 or y == 4:
state_features[5*x+y] = np.array([1,0,0])
state_features[5*2+2] = np.array([0,0,0])
state_features[5*2+4] = np.array([1,0,1])
modelA = IRLModel(25, 4, 2, 3, T, 0.95, 0.01, state_features)
modelA_IRL = IRLModel(25, 4, 2, 3, T, 0.95, 0.01, state_features)
modelA_EM = IRLModel(25, 4, 2, 3, T, 0.95, 0.01, state_features, ignore_tau=True)
modelA_EM.set_tau(tau_EM())
nu = np.zeros(25)
nu[5*2+0] = 1.0
sigma = np.array([0.5, 0.5])
Theta = np.array([[0, 20, 30],
[20, 0, 30]])
tau = np.zeros((2, 2, 25))
for i in xrange(25):
tau[0][0][i] = tau[1][1][i] = 1.0
for i in [(5*2), (5*2+2)]:
tau[0][1][i] = 0.7
tau[0][0][i] = 0.3
tau[1][0][i] = tau[1][1][i] = 0.5
simA = Simulator(modelA, nu, T, sigma, Theta, tau=tau)
trajectories = simA.trajectories(125, 20)
trajectories_test = simA.trajectories(40, 20)
irl, mlirl, expert = [], [], []
for batch in xrange(1,6):
modelA_IRL = IRLModel(25, 4, 2, 3, T, 0.95, 0.01, state_features)
modelA_EM = IRLModel(25, 4, 2, 3, T, 0.95, 0.01, state_features, ignore_tau=True)
modelA_EM.set_tau(tau_EM())
modelA_IRL.learn(trajectories[:batch*25], 1e-3, 7)
print "Us"
irl.append(modelA_IRL.test(trajectories_test))
print "MLIRL"
modelA_EM.learn(trajectories[:batch*25], 1e-3, 7)
mlirl.append(modelA_EM.test(trajectories_test))
print "Expert"
expert.append(modelA.test(trajectories_test))
print irl, mlirl, expert
def worldB():
T = getT()
state_features = np.zeros((25, 2))
dynamic_features = np.zeros((25,2))
for x in xrange(5):
for y in xrange(5):
state_features[5*x + y] = np.array([0,0])
dynamic_features[5*x+ y] = np.array([0,1])
state_features[5*4+0] = np.array([0,1])
state_features[5*0+4] = np.array([1,0])
dynamic_features[5*2+0] = np.array([1,1])
dynamic_features[5*2+1] = np.array([1,1])
dynamic_features[5*2+2] = np.array([1,1])
dynamic_features[5*3+2] = np.array([1,1])
dynamic_features[5*4+2] = np.array([1,1])
modelB = IRLModel(25, 4, 2, 2, T, 0.95, 0.1, state_features, dynamic_features=dynamic_features)
modelB_IRL = IRLModel(25, 4, 2, 2, T, 0.95, 0.1, state_features,dynamic_features=dynamic_features)
modelB_EM = IRLModel(25, 4, 2, 2, T, 0.95, 0.1, state_features)
modelB_EM.set_tau(tau_EM())
nu = np.zeros(25)
nu[5*2+0] = 1.0
sigma = np.array([0.0, 1.0])
Theta = np.array([[30,0],
[0, 30]])
omega = np.zeros([2,2,2])
omega[0,0] = np.array([-11,12])
omega[1,0] = np.array([13,-12])
simB = Simulator(modelB, nu, T, sigma, Theta, omega=omega)
trajectories = simB.trajectories(125, 20)
trajectories_test = simB.trajectories(40, 20)
irl, mlirl, expert = [], [], []
for batch in xrange(1,6):
modelB_IRL = IRLModel(25, 4, 2, 2, T, 0.95, 0.01, state_features,dynamic_features=dynamic_features)
modelB_EM = IRLModel(25, 4, 2, 2, T, 0.95, 0.01, state_features,ignore_tau=True)
modelB_EM.set_tau(tau_EM())
modelB_IRL.learn(trajectories[:125], 1e-3, 7)
irl.append(modelB_IRL.test(trajectories_test))
modelB_EM.learn(trajectories[:batch*25], 1e-3, 7)
mlirl.append(modelB_EM.test(trajectories_test))
expert.append(modelB.test(trajectories_test))
print irl, mlirl, expert
