Solving OpenAI's Cartpole with a very simple PID controller in 35 lines

pid_cartpole.py

import numpy as np
import gym

def sigmoid(x):
    return 1.0 / (1.0 + np.exp(-x))

env = gym.make('CartPole-v1')
desired_state = np.array([0, 0, 0, 0])
desired_mask = np.array([0, 0, 1, 0])

P, I, D = 0.1, 0.01, 0.5

for i_episode in range(20):
    state = env.reset()
    integral = 0
    derivative = 0
    prev_error = 0
    for t in range(500):
        env.render()
        error = state - desired_state

        integral += error
        derivative = error - prev_error
        prev_error = error

        pid = np.dot(P * error + I * integral + D * derivative, desired_mask)
        action = sigmoid(pid)
        action = np.round(action).astype(np.int32)

        state, reward, done, info = env.step(action)
        if done:
            print("Episode finished after {} timesteps".format(t+1))
            break
env.close()

Hi! I am trying to figure out why the Pole Angle error is the only one you care about (desired_mask = np.array([0,0,1,0]). Was this trial and error or is there a specific reason why that error value is more important? Thanks!

desired_mask = np.array([1, 1, 1, 1])

P, I, D = [1/150, 1/950, 0.1, 0.01], [0.0005, 0.001, 0.01, 0.0001], [0.2, 0.0001, 0.5, 0.005]

with this change you can see how every state's component is taken into calculations
also try this following state = env.reset()

env.state[0] *= 30

this snippet disable fps limit

import pyglet
pyglet.options['vsync'] = False