added dubins car model

control_affine_models/dubins_car.py

@@ -0,0 +1,147 @@
+import warnings
+from contextlib import contextmanager
+from copy import copy
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+from scipy.integrate import solve_ivp
+from scipy.linalg import LinAlgWarning
+from sklearn.exceptions import ConvergenceWarning
+from sklearn.linear_model import Lasso
+import pickle
+import dill
+import pysindy as ps
+
+from utils import *
+
+# Seed the random number generators for reproducibility
+#np.random.seed(100)
+
+# Set up simulation parameters
+time_horzn = 1.0
+dt = 0.01
+ang_ind = [2] # theta_
+
+# Initialize integrator keywords for solve_ivp to replicate the odeint defaults
+integrator_keywords = {}
+integrator_keywords["rtol"] = 1e-9
+integrator_keywords["method"] = "LSODA"
+integrator_keywords["atol"] = 1e-9
+
+# Randomized initial condition
+def x0_fun(): return [0.0, 0.0, np.random.uniform(-np.pi, np.pi)]
+
+# Range of the amplitudes and frequencies of the randomized sine inputs
+u_amp_range = [0, np.pi]
+u_freq_range = [0, 5]
+
+# Model parameter
+v = 1.0
+
+def dubins_car_dyn(state, u):
+    # x_dot = f(x, u)
+
+    x_, y_, theta_ = state
+
+    x_dot = v * np.cos(theta_)
+    y_dot = v * np.sin(theta_)
+    theta_dot = u
+
+    return [x_dot, y_dot, theta_dot]
+
+def dubins_car(t, state, u_fun):
+    u = u_fun(t)
+    return dubins_car_dyn(state, u)
+
+## Train a SINDYc model using trajectory data
+# Generate the training dataset
+t_data = np.arange(0, time_horzn, dt)
+t_data_span = (t_data[0], t_data[-1])
+n_traj_train = 5000
+
+x_train, x_dot_train, u_train = gen_trajectory_dataset(dubins_car, x0_fun, n_traj_train, time_horzn, dt, 
+                                          u_amp_range, u_freq_range, ang_ind, **integrator_keywords)
+
+#plt.plot(t_data, x_train[0])
+#plt.show()
+#plt.plot(t_data, x_train[0][:,1], t_data, x_dot_train[0][:,0])
+#plt.show()
+#plt.plot(t_data, x_train[0][:,3], t_data, x_dot_train[0][:,2])
+#plt.show()
+#plt.plot(t_data, u_train[0])
+#plt.show()
+
+# Instantiate and fit the SINDYc model
+# Generalized Library (such that it's control affine)
+generalized_library = ps.GeneralizedLibrary(
+    [ps.PolynomialLibrary(degree = 20),
+     #ps.FourierLibrary(n_frequencies = 1),
+     ps.IdentityLibrary() # for control input
+    ],
+    #tensor_array = [[0,1,0]],
+    inputs_per_library = [[2], [3]]
+)
+
+# Unconstrained model
+model_uc = ps.SINDy(
+    optimizer = ps.STLSQ(threshold = 0.0001),
+    feature_library = generalized_library,
+)
+model_uc.fit(x_train, x_dot = x_dot_train, u = u_train, t = dt)
+model_uc.print()
+print("Feature names:\n", model_uc.get_feature_names())
+
+model = model_uc
+
+control_affine = check_control_affine(model)
+assert control_affine is True
+
+## Assess results on a test trajectory
+# Evolve the equations in time using a different initial condition
+x0 = x0_fun()
+u_amp_data = np.random.uniform(0, 100)
+u_freq_data = np.random.uniform(0, 5)
+u_fun = lambda t: u_amp_data * np.sin(2 * np.pi * u_freq_data * t)
+test_model_prediction(dubins_car, model, x0, u_fun, time_horzn, dt, ang_ind, **integrator_keywords)
+u_amp_data = np.random.uniform(0, 100)
+u_freq_data = np.random.uniform(0, 5)
+test_model_prediction(dubins_car, model, x0, u_fun, time_horzn, dt, ang_ind, **integrator_keywords)
+u_amp_data = np.random.uniform(0, 100)
+u_freq_data = np.random.uniform(0, 5)
+test_model_prediction(dubins_car, model, x0, u_fun, time_horzn, dt, ang_ind, **integrator_keywords)
+
+## Compute conformal prediction quantile
+x_max = 10.0
+y_max = 10.0
+theta_max = np.pi
+
+x_range = np.array([
+     [-x_max, x_max],
+     [-y_max, y_max],
+     [-theta_max, theta_max]
+])
+
+u_range = np.array([
+     [-np.pi, np.pi]
+])
+
+alpha = 0.05
+n_cal = 1000
+n_val = 1000
+norm = 2
+
+quantile = get_conformal_prediction_quantile(dubins_car_dyn, model, x_range, u_range,
+                                      n_cal, n_val, alpha, norm)
+
+# Save the quantile and alpha as paramters under the model
+model_error = {"alpha": alpha, "quantile": quantile, "norm": norm}
+model.model_error = model_error
+
+## Save the model and dataset
+with open('./control_affine_models/saved_models/model_dubins_car_sindy', 'wb') as file:
+    dill.dump(model, file)
+
+# Testing
+with open('./control_affine_models/saved_models/' + 'model_dubins_car_sindy', 'rb') as file:
+	model2 = dill.load(file)