Last active
February 6, 2020 06:08
-
-
Save flyman3046/d37680eeaac469a4030c690ae65b0419 to your computer and use it in GitHub Desktop.
Implementation of Evolution Strategies to Solve CartPole-v0
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# https://gist.github.com/karpathy/77fbb6a8dac5395f1b73e7a89300318d | |
import gym | |
import numpy as np | |
def f(env, weight): | |
total_reward = 0.0 | |
num_run = 100 | |
for t in range(num_run): | |
observation = env.reset() | |
for i in range(300): | |
action = 1 if np.dot(weight, observation) > 0 else 0 | |
observation, reward, done, info = env.step(action) | |
total_reward += reward | |
if done: | |
break | |
return total_reward / num_run | |
def evolution_strategy(env): | |
# hyperparameters | |
npop = 50 # population size | |
sigma = 0.1 # noise standard deviation | |
alpha = 0.001 # learning rate | |
# start the optimization | |
weight = np.random.rand(4) # our initial guess is random | |
for i in range(50): | |
# print current fitness of the most likely parameter setting | |
print 'iter {}. weight: {}, reward: {}'.format(i, str(weight), f(env, weight)) | |
# initialize memory for a population of w's, and their rewards | |
N = np.random.randn(npop, 4) # samples from a normal distribution N(0,1) | |
R = np.zeros(npop) | |
for j in range(npop): | |
w_try = weight + sigma * N[j] # jitter w using gaussian of sigma 0.1 | |
R[j] = f(env, w_try) # evaluate the jittered version | |
# standardize the rewards to have a gaussian distribution | |
A = (R - np.mean(R)) / np.std(R) | |
# perform the parameter update. The matrix multiply below | |
# is just an efficient way to sum up all the rows of the noise matrix N, | |
# where each row N[j] is weighted by A[j] | |
weight = weight + alpha / (npop * sigma) * np.dot(N.T, A) | |
env = gym.make('CartPole-v0') | |
evolution_strategy(env) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment