def make_env(env_name, seed=-1, render_mode=False):
if (env_name.startswith("CartPoleSwingUp")):
if (env_name.startswith("CartPoleSwingUpHarder")):
print("cartpole_swingup_harder_started")
from cartpole_swingup_harder import CartPoleSwingUpHarderEnv
env = CartPoleSwingUpHarderEnv()
else:
print("cartpole_swingup_started")
from cartpole_swingup import CartPoleSwingUpEnv
env = CartPoleSwingUpEnv()
elif (env_name.startswith("DreamCartPoleSwingUp")):
print("dream_cartpole_swingup_started")
from dream import DreamCartPoleSwingUpEnv
env = DreamCartPoleSwingUpEnv()
else:
assert False, "invalid environment name."
if (seed >= 0):
env.seed(seed)
'''
print("environment details")
print("env.action_space", env.action_space)
print("high, low", env.action_space.high, env.action_space.low)
print("environment details")
print("env.observation_space", env.observation_space)
print("high, low", env.observation_space.high, env.observation_space.low)
assert False
'''
return env
return Model(5, 1, 16)
def _evaluate_once(self, env, model):
obs = env.reset()
h = model.M.reset()
rewards = 0
done = False
while not done:
action = model(obs, h, module="C")
obs, reward, done, _ = env.step(action)
h = model(obs, action, module="M")
rewards += reward
return rewards
env = CartPoleSwingUpEnv()
es = OpenaiES(len(Model(5, 1, 16)), sigma=args.sigma, stepsize=args.stepsize)
global_best_fitness = -np.inf
with CartPoleSwingUpEvaluator(args.num_workers, args.models_per_worker,
args.precision) as evaluator:
popsize = len(evaluator)
for gen in range(args.num_gen):
seeds, solutions = es.sample(popsize)
fitness, success = evaluator.evaluate(seeds, solutions, args.num_evals)
assert success, f"evaluation failed at generation {gen}"
es.step(fitness)
best_fitness = np.max(fitness)
if best_fitness > global_best_fitness:
print(
def _build_env(self):
return CartPoleSwingUpEnv()
def make_env(name):
"""Simple helper function to load CartPoleSwingUp or a gym environment."""
if name == 'CartPoleSwingUp':
return CartPoleSwingUpEnv()
else:
return gym.make(name)