def cartpole_sampling(theta, cm, K, Ke, N, epsilon):
theta_list = np.random.multivariate_normal(theta, cm, K)
result_list = []
for x in range(K):
# concurrent_eval(theta_list, x, result_list, N)
avg_reward = 0
for i in range(N):
cartpole = CartPole()
cartpole.pi_params = theta_list[x].reshape(4, 2)
epi = CartPoleEpisode(cartpole)
avg_reward += epi.run_all_steps()
result_list.append((theta_list[x], avg_reward / N))
# print(sorted(result_list, key=lambda n: n[-1], reverse=True))
elite_list = sorted(result_list, key=lambda n: n[-1], reverse=True)[:Ke]
# print(elite_list)
theta_final = np.zeros(8)
cm_final = epsilon * np.identity(8)
J_final = 0
for t in elite_list:
theta_final += t[0]
cm_final += np.array([t[0] - theta]).T.dot(np.array([t[0] - theta]))
J_final += t[1]
theta_final /= Ke
cm_final /= (epsilon + Ke)
# print(cm_final)
J_final /= Ke
return theta_final, cm_final, J_final
def cartpole_evaluate(table, N):
avg_reward = 0
for i in range(N):
cartpole = CartPole()
cartpole.pi_params = table
epi = CartPoleEpisode(cartpole)
avg_reward += epi.run_all_steps()
return avg_reward / N
def multi_cartpole_episode(table, l):
for i in l:
cartpole = CartPole()
# print(i)
cartpole.pi_params = table
epi = CartPoleEpisode(cartpole)
cp_q.put(epi.run_all_steps())
return 0
def cartpole_evaluate(t, N):
reward_l = []
for i in range(N):
cartpole = CartPole()
# print(i)
cartpole.pi_params = t.reshape(4, 2)
epi = CartPoleEpisode(cartpole)
reward_l.append(epi.run_all_steps())
return sum(reward_l) / N