UH-CMA-ES 10.0

UH-CMA-ES10.py

import argparse
import numpy
import keras
import gym
import cma

parser = argparse.ArgumentParser()
parser.add_argument("environment")
args = parser.parse_args()

environment = gym.make(args.environment)

model = keras.models.Sequential([
	keras.layers.Dense(10,	activation="tanh", input_shape=environment.observation_space.shape),
	keras.layers.Dense(5,	activation="tanh"),
	keras.layers.Dense(environment.action_space.n)])

shapes = [weight.shape for weight in model.get_weights()]

def get_solution(weights):
	return numpy.concatenate([weight.reshape(-1) for weight in weights])

def set_weights(solution):
	model.set_weights([solution[1:1+numpy.prod(shape)].reshape(shape) for shape in shapes])

def get_action(observation):
	return numpy.argmax(model.predict_on_batch(observation))

shape = (1,) + environment.observation_space.shape

def get_reward():
	observation = environment.reset()
	Reward = 0
	done = False
	while not done:
		observation = observation.reshape(shape)
		action = get_action(observation)
		observation, reward, done, _info = environment.step(action)
		Reward += reward
	return Reward

def f(x):
	set_weights(x)
	Reward = get_reward()
	return -Reward

x0 = get_solution(model.get_weights())

options = {}

options['maxfevals']	= 1e4

options['tolx']		= 0
options['tolfun']	= 0
options['tolfunhist']	= 0

noise_handler = cma.NoiseHandler(len(x0))

environment.monitor.start("gym", video_callable=False)

cma.fmin(f, x0, 10.0, options, noise_handler=noise_handler)

environment.monitor.close()

gym.upload("gym", algorithm_id="alg_WPVlCpgIRfC31ilRO1moQ")