def fit_cartpole():
global _env
hist = []
_env = gym.make('CartPole-v1')
_env._max_episode_steps = 500
nn = SimpleNeuralControllerNumpy(4, 1, 2, 5)
nn.init_random_params()
es = cma.CMAEvolutionStrategy(nn.get_parameters(), 0.2)
for _ in range(800):
solutions = es.ask()
es.tell(solutions, [eval_nn(x) for x in solutions])
hist.append(-es.result.fbest)
_env.close()
return hist, -es.result.fbest, es.result.xbest
action = 1
else:
action = 0
observation, reward, done, info = env.step(action)
total_reward += reward
if done:
print("Episode finished after %d timesteps" % (t + 1))
break
return -total_reward
sigma = 1
### A completer pour optimiser les parametres du reseau de neurones avec CMA-ES ###
nn = SimpleNeuralControllerNumpy(4, 1, 2, 5)
nn.init_random_params()
res = launch_cmaes_full_genotype(nn.get_parameters(),
sigma,
nbeval=1000,
display=True,
ma_func=eval_nn)
nn.set_parameters(res)
env.reset()
r = env.step(env.action_space.sample()) # take a random action
observations = r[0]
reward = r[1]
done = r[2]
print(nn.predict(observations))
for _ in range(1000):