def test_rendering(self):
bench = CMAESBenchmark()
env = bench.get_environment()
wrapped = StateTrackingWrapper(env)
wrapped.reset()
with pytest.raises(NotImplementedError):
wrapped.render_state_tracking()
bench = CMAESBenchmark()
def dummy():
return [1, [2, 3]]
bench.config.state_method = dummy
bench.config.observation_space = gym.spaces.Tuple(
(
gym.spaces.Discrete(2),
gym.spaces.Box(low=np.array([-1, 1]), high=np.array([5, 5])),
)
)
env = bench.get_environment()
wrapped = StateTrackingWrapper(env)
wrapped.reset()
with pytest.raises(NotImplementedError):
wrapped.render_state_tracking()
def dummy2():
return [0.5]
bench.config.state_method = dummy2
bench.config.observation_space = gym.spaces.Box(
low=np.array([0]), high=np.array([1])
)
env = bench.get_environment()
wrapped = StateTrackingWrapper(env)
wrapped.reset()
wrapped.step(1)
wrapped.step(1)
img = wrapped.render_state_tracking()
self.assertTrue(img.shape[-1] == 3)
bench = LubyBenchmark()
env = bench.get_environment()
wrapped = StateTrackingWrapper(env, 2)
wrapped.reset()
wrapped.step(1)
wrapped.step(1)
img = wrapped.render_state_tracking()
self.assertTrue(img.shape[-1] == 3)
class discrete_obs_env:
def __init__(self):
self.observation_space = gym.spaces.Discrete(2)
self.action_space = gym.spaces.Discrete(2)
self.reward_range = (1, 2)
self.metadata = {}
def reset(self):
return 1
def step(self, action):
return 1, 1, 1, 1
env = discrete_obs_env()
wrapped = StateTrackingWrapper(env, 2)
wrapped.reset()
wrapped.step(1)
img = wrapped.render_state_tracking()
self.assertTrue(img.shape[-1] == 3)
class multi_discrete_obs_env:
def __init__(self):
self.observation_space = gym.spaces.MultiDiscrete([2, 3])
self.action_space = gym.spaces.Discrete(2)
self.reward_range = (1, 2)
self.metadata = {}
def reset(self):
return [1, 2]
def step(self, action):
return [1, 2], 1, 1, 1
env = multi_discrete_obs_env()
wrapped = StateTrackingWrapper(env)
wrapped.reset()
wrapped.step(1)
img = wrapped.render_state_tracking()
self.assertTrue(img.shape[-1] == 3)
class multi_binary_obs_env:
def __init__(self):
self.observation_space = gym.spaces.MultiBinary(2)
self.action_space = gym.spaces.Discrete(2)
self.reward_range = (1, 2)
self.metadata = {}
def reset(self):
return [1, 1]
def step(self, action):
return [1, 1], 1, 1, 1
env = multi_binary_obs_env()
wrapped = StateTrackingWrapper(env)
wrapped.reset()
wrapped.step(1)
img = wrapped.render_state_tracking()
self.assertTrue(img.shape[-1] == 3)
def test_get_env(self):
bench = CMAESBenchmark()
env = bench.get_environment()
self.assertTrue(issubclass(type(env), CMAESEnv))
import numpy as np
from examples.example_utils import make_chainer_a3c, train_chainer
from dacbench.benchmarks import CMAESBenchmark
from dacbench.wrappers import EpisodeTimeWrapper
# Helper function to flatten observation space
def flatten(li):
return [value for sublist in li for value in sublist]
# Make CMA-ES environment
# We use the configuration from the "Learning to Optimize Step-size Adaption in CMA-ES" Paper by Shala et al.
bench = CMAESBenchmark()
env = bench.get_environment()
# Wrap environment to track time
# Here we also want the mean of each 2 step interval
env = EpisodeTimeWrapper(env, 2)
# Make chainer agent
space_array = [
env.observation_space[k].low
for k in list(env.observation_space.spaces.keys())
]
obs_size = np.array(flatten(space_array)).size
action_size = env.action_space.low.size
agent = make_chainer_a3c(obs_size, action_size)
# Train agent for 10 episodes