2023-02-02 17:54:09 +00:00
|
|
|
import numpy as np
|
|
|
|
from sklearn.gaussian_process import GaussianProcessRegressor
|
|
|
|
from sklearn.gaussian_process.kernels import Matern
|
|
|
|
|
|
|
|
from PolicyModel.GaussianModel import GaussianPolicy
|
|
|
|
from AcquistionFunctions.ExpectedImprovement import ExpectedImprovement
|
2023-02-06 14:43:30 +00:00
|
|
|
from AcquistionFunctions.ProbabilityOfImprovement import ProbabilityOfImprovement
|
|
|
|
from AcquistionFunctions.ConfidenceBound import ConfidenceBound
|
2023-02-02 17:54:09 +00:00
|
|
|
|
2023-02-03 13:05:03 +00:00
|
|
|
from ToyTask.MountainCarGym import Continuous_MountainCarEnv
|
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
import gym
|
|
|
|
import time
|
|
|
|
|
|
|
|
import matplotlib.pyplot as plt
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
class BayesianOptimization:
|
2023-02-03 13:05:03 +00:00
|
|
|
def __init__(self, env, nr_step, nr_init=3, acq='ei', nr_weights=6, policy_seed=None):
|
2023-02-02 17:54:09 +00:00
|
|
|
self.env = env
|
|
|
|
self.nr_init = nr_init
|
|
|
|
self.acq = acq
|
|
|
|
self.X = None
|
|
|
|
self.Y = None
|
2023-02-06 14:43:30 +00:00
|
|
|
self.counter_array = np.zeros((1, 1))
|
2023-02-02 17:54:09 +00:00
|
|
|
self.gp = None
|
|
|
|
|
|
|
|
self.episode = 0
|
|
|
|
self.best_reward = np.empty((1, 1))
|
2023-02-06 14:43:30 +00:00
|
|
|
self.distance_penalty = 0
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
self.nr_policy_weights = nr_weights
|
|
|
|
self.nr_steps = nr_step
|
|
|
|
self.policy_seed = policy_seed
|
|
|
|
self.lowerb = -1.0
|
|
|
|
self.upperb = 1.0
|
|
|
|
self.policy_model = GaussianPolicy(self.nr_policy_weights,
|
|
|
|
self.nr_steps,
|
|
|
|
self.policy_seed,
|
|
|
|
self.lowerb,
|
|
|
|
self.upperb)
|
|
|
|
|
|
|
|
self.nr_test = 100
|
|
|
|
|
|
|
|
def initialize(self):
|
|
|
|
self.env.reset()
|
2023-02-06 14:43:30 +00:00
|
|
|
if self.env.render_mode == 'human':
|
|
|
|
self.env.render()
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
self.X = np.zeros((self.nr_init, self.nr_policy_weights))
|
|
|
|
self.Y = np.zeros((self.nr_init, 1))
|
|
|
|
|
|
|
|
for i in range(self.nr_init):
|
|
|
|
self.policy_model.random_policy()
|
|
|
|
self.X[i, :] = self.policy_model.weights.T
|
|
|
|
policy = self.policy_model.policy_rollout()
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
reward, step_count = self.runner(policy)
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
self.Y[i] = reward
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
self.gp = GaussianProcessRegressor(Matern(nu=1.5))
|
|
|
|
self.gp.fit(self.X, self.Y)
|
|
|
|
|
|
|
|
def runner(self, policy):
|
|
|
|
done = False
|
|
|
|
step_count = 0
|
|
|
|
env_reward = 0.0
|
|
|
|
while not done:
|
|
|
|
action = policy[step_count]
|
|
|
|
output = self.env.step(action)
|
|
|
|
env_reward += output[1]
|
|
|
|
done = output[2]
|
2023-02-06 14:43:30 +00:00
|
|
|
if self.env.render_mode == 'human':
|
|
|
|
time.sleep(0.0001)
|
2023-02-02 17:54:09 +00:00
|
|
|
step_count += 1
|
|
|
|
if step_count >= self.nr_steps:
|
|
|
|
done = True
|
|
|
|
distance = -(self.env.goal_position - output[0][0])
|
|
|
|
env_reward += distance * self.distance_penalty
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
if self.counter_array[0] == 0:
|
|
|
|
|
|
|
|
if step_count == 100:
|
|
|
|
step_count = np.NAN
|
|
|
|
self.counter_array[0] = step_count
|
|
|
|
|
|
|
|
else:
|
|
|
|
if step_count == 100:
|
|
|
|
step_count = np.NAN
|
|
|
|
self.counter_array = np.vstack((self.counter_array, step_count))
|
|
|
|
|
|
|
|
if self.env.render_mode == 'human':
|
|
|
|
time.sleep(0.25)
|
2023-02-02 17:54:09 +00:00
|
|
|
self.env.reset()
|
2023-02-06 14:43:30 +00:00
|
|
|
return env_reward, step_count
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
def next_observation(self):
|
|
|
|
if self.acq == 'ei':
|
|
|
|
x_next = ExpectedImprovement(self.gp,
|
|
|
|
self.X,
|
|
|
|
self.nr_test,
|
|
|
|
self.nr_policy_weights,
|
2023-02-06 14:43:30 +00:00
|
|
|
kappa=0,
|
2023-02-02 17:54:09 +00:00
|
|
|
seed=self.policy_seed,
|
|
|
|
lower=self.lowerb,
|
|
|
|
upper=self.upperb)
|
|
|
|
|
|
|
|
return x_next
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
elif self.acq == 'pi':
|
|
|
|
x_next = ProbabilityOfImprovement(self.gp,
|
|
|
|
self.X,
|
|
|
|
self.nr_test,
|
|
|
|
self.nr_policy_weights,
|
|
|
|
kappa=2.576,
|
|
|
|
seed=self.policy_seed,
|
|
|
|
lower=self.lowerb,
|
|
|
|
upper=self.upperb)
|
|
|
|
|
|
|
|
return x_next
|
|
|
|
|
|
|
|
elif self.acq == 'cb':
|
|
|
|
x_next = ConfidenceBound(self.gp,
|
|
|
|
self.X,
|
|
|
|
self.nr_test,
|
|
|
|
self.nr_policy_weights,
|
|
|
|
lam=2.576,
|
|
|
|
seed=self.policy_seed,
|
|
|
|
lower=self.lowerb,
|
|
|
|
upper=self.upperb)
|
|
|
|
|
|
|
|
return x_next
|
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
else:
|
|
|
|
raise NotImplementedError
|
|
|
|
|
|
|
|
def eval_new_observation(self, x_next):
|
|
|
|
self.policy_model.weights = x_next
|
|
|
|
policy = self.policy_model.policy_rollout()
|
2023-02-06 14:43:30 +00:00
|
|
|
reward, step_count = self.runner(policy)
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
self.X = np.vstack((self.X, x_next))
|
|
|
|
self.Y = np.vstack((self.Y, reward))
|
|
|
|
|
|
|
|
self.gp.fit(self.X, self.Y)
|
|
|
|
|
|
|
|
if self.episode == 0:
|
|
|
|
self.best_reward[0] = max(self.Y)
|
|
|
|
else:
|
|
|
|
self.best_reward = np.vstack((self.best_reward, max(self.Y)))
|
|
|
|
|
|
|
|
self.episode += 1
|
2023-02-06 14:43:30 +00:00
|
|
|
return step_count
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
def plot_reward(self):
|
|
|
|
epsiodes = np.linspace(0, self.episode, self.episode)
|
|
|
|
plt.plot(epsiodes, self.best_reward)
|
2023-02-06 14:43:30 +00:00
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
plt.show()
|
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
def plot_objective_function(self):
|
|
|
|
plt.scatter(self.X[:, 0], self.Y)
|
|
|
|
|
|
|
|
x_dom = np.linspace(self.lowerb * np.ones((1, 6)), self.upperb * np.ones((1, 6)), 100).reshape(-1, self.nr_policy_weights)
|
|
|
|
y_plot, y_sigma = self.gp.predict(x_dom, return_std=True)
|
|
|
|
|
|
|
|
y_plot = y_plot.reshape(-1)
|
|
|
|
y_sigma = y_sigma.reshape(-1)
|
|
|
|
print(y_plot.shape, x_dom.shape)
|
|
|
|
|
|
|
|
plt.plot(x_dom[:, 0], y_plot)
|
|
|
|
plt.fill_between(
|
|
|
|
x_dom[:, 0],
|
|
|
|
y_plot - 1.96 * y_sigma,
|
|
|
|
y_plot + 1.96 * y_sigma,
|
|
|
|
alpha=0.5
|
|
|
|
)
|
|
|
|
plt.show()
|
|
|
|
|
|
|
|
def get_best_result(self):
|
|
|
|
y_hat = self.gp.predict(self.X)
|
|
|
|
idx = np.argmax(y_hat)
|
|
|
|
x_max = self.X[idx, :]
|
|
|
|
self.policy_model.weights = x_max
|
|
|
|
self.policy_model.policy_rollout()
|
|
|
|
print(self.counter_array[idx], idx)
|
|
|
|
self.policy_model.plot_policy(finished=self.counter_array[idx])
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
def main():
|
2023-02-03 13:05:03 +00:00
|
|
|
nr_steps = 100
|
2023-02-06 14:43:30 +00:00
|
|
|
env = Continuous_MountainCarEnv() # render_mode='human'
|
|
|
|
bo = BayesianOptimization(env, nr_steps, acq='cb')
|
2023-02-02 17:54:09 +00:00
|
|
|
bo.initialize()
|
2023-02-03 13:05:03 +00:00
|
|
|
iteration_steps = 200
|
2023-02-02 17:54:09 +00:00
|
|
|
for i in range(iteration_steps):
|
|
|
|
x_next = bo.next_observation()
|
2023-02-06 14:43:30 +00:00
|
|
|
step_count = bo.eval_new_observation(x_next)
|
2023-02-02 17:54:09 +00:00
|
|
|
|
2023-02-06 14:43:30 +00:00
|
|
|
print(bo.episode, bo.best_reward[-1][0], step_count)
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
bo.plot_reward()
|
2023-02-06 14:43:30 +00:00
|
|
|
bo.get_best_result()
|
|
|
|
plt.show()
|
|
|
|
|
2023-02-02 17:54:09 +00:00
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
main()
|