def objective_two(
individual: cgp.IndividualSingleGenome,
cum_reward_threshold: int,
gamma: float,
seed: int,
rng: np.random.Generator,
):
if not individual.fitness_is_None():
return individual
# environment initialization
env = gym.make('CartPole-v0')
try:
individual.fitness = inner_objective(
individual,
network=individual.network,
env=env,
cum_reward_threshold=cum_reward_threshold,
gamma=gamma,
seed=seed,
rng=rng,
mode='episode_min')
except ZeroDivisionError:
individual.fitness = -np.inf
return individual
def _create_individual(genome, fitness=None, individual_type="SingleGenome"):
if individual_type == "SingleGenome":
ind = IndividualSingleGenome(genome)
if fitness is not None:
ind.fitness = fitness
return ind
elif individual_type == "MultiGenome":
ind = IndividualMultiGenome([genome])
if fitness is not None:
ind.fitness = fitness
return ind
else:
raise NotImplementedError("Unknown individual type.")
def objective(
individual: cgp.IndividualSingleGenome,
network_params: dict,
curriculum_params: dict,
seeds
):
if not individual.fitness_is_None():
return individual
try:
individual.fitness = inner_objective(individual, network_params, curriculum_params, seeds)
except ZeroDivisionError:
individual.fitness = -np.inf
return individual
def objective_one(
individual: cgp.IndividualSingleGenome,
n_episodes: int,
gamma: float,
seed: int,
rng: np.random.Generator,
):
if not individual.fitness_is_None():
return individual
# environment initialization
env = gym.make('CartPole-v0')
env.seed(seed=seed)
env.action_space.seed(seed)
# network initialization
torch.manual_seed(seed=seed)
network = Network(n_inputs=env.observation_space.shape[0],
n_hidden=100,
n_outputs=env.action_space.n,
learning_rate=2e-4,
weight_update_mode='evolved_rule')
try:
individual.fitness = inner_objective(individual,
network=network,
env=env,
n_episodes=n_episodes,
gamma=gamma,
seed=seed,
rng=rng,
mode='reward_max')
except ZeroDivisionError:
individual.fitness = -np.inf
# Todo write network.state_dict() to ind (and possibly pickle dumps)
individual.network = network # assign the trained network to the individual for objective 2
return individual