Random Controllers for OpenAI Gym

simple_pend.py

import gym
import numpy as np
import random

env = gym.make('Pendulum-v0')
dim = env.observation_space.shape[0] + 1
params = int(dim + (dim*(dim-1))/2)

# linear controller with pairwise features
def quad_control(w,ob,t):
    obs = list(ob) + [1]
    for i in range(dim):
        for j in range(i+1,dim):
            obs.append(obs[i]*obs[j])
    s = w * np.array(obs)
    return np.clip([s.sum()],-2.0,2.0)

# run an episode
def func(x):
    errs = []
    for i_episode in range(8):
        err = 0.0
        observation = env.reset()
        for t in range(200):
            action = quad_control(x,observation,t)
            observation, reward, done, info = env.step(action)
            err -= reward
            if done: # early exit
                break
        errs.append(err)
    return sorted(errs)[6]

# try 1000 random configurations
best_score = float('inf')
for _ in range(1000):
    x = 10.0*(np.random.random(params)*2.0 - 1.0)
    s = func(x)
    if s < best_score:
        best_score = s
        resx = x
        print('{}\t{:.2f}'.format(_,s))
    if s < 500:
        print('early exit!')
        break
print(_,resx)

# see if it worked
for i_episode in range(10):
    observation = env.reset()
    err = 0.0
    for t in range(200):
        env.render()
        action = quad_control(resx,observation,t)
        observation, reward, done, info = env.step(action)
        err -= reward
        if done:
            print("Episode finished after {} timesteps".format(t+1))
            break
    print('{:.2f}'.format(err))
print("we're done!")