def setUp(self) -> None:
self.temp_dir = tempfile.TemporaryDirectory()
episodes = 80
seeds = [0, 1, 3, 4, 5]
experiment_name = "test_env"
logger = Logger(
output_path=Path(self.temp_dir.name),
experiment_name=experiment_name,
step_write_frequency=None,
episode_write_frequency=None,
)
benchmark = SigmoidBenchmark()
env = benchmark.get_benchmark()
agent = RandomAgent(env)
logger.set_env(env)
env_logger = logger.add_module(env)
for seed in seeds:
env.seed(seed)
logger.set_additional_info(seed=seed)
logger.reset_episode()
for episode in range(episodes):
state = env.reset()
done = False
reward = 0
step = 0
while not done:
action = agent.act(state, reward)
env_logger.log(
"logged_step",
step,
)
env_logger.log(
"logged_episode",
episode,
)
next_state, reward, done, _ = env.step(action)
env_logger.log(
"reward",
reward,
)
env_logger.log(
"done",
done,
)
agent.train(next_state, reward)
state = next_state
logger.next_step()
step += 1
agent.end_episode(state, reward)
logger.next_episode()
env.close()
logger.close()
self.log_file = env_logger.log_file.name
def run_policy(results_path,
benchmark_name,
num_episodes,
policy,
seeds=np.arange(10)):
bench = getattr(benchmarks, benchmark_name)()
for s in seeds:
if benchmark_name == "CMAESBenchmark":
experiment_name = f"csa_{s}"
else:
experiment_name = f"optimal_{s}"
logger = Logger(experiment_name=experiment_name,
output_path=results_path / benchmark_name)
env = bench.get_benchmark(seed=s)
env = PerformanceTrackingWrapper(
env, logger=logger.add_module(PerformanceTrackingWrapper))
agent = GenericAgent(env, policy)
logger.add_agent(agent)
logger.add_benchmark(bench)
logger.set_env(env)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def run_dacbench(results_path,
agent_method,
num_episodes,
bench=None,
seeds=None):
"""
Run all benchmarks for 10 seeds for a given number of episodes with a given agent and save result
Parameters
-------
bench
results_path : str
Path to where results should be saved
agent_method : function
Method that takes an env as input and returns an agent
num_episodes : int
Number of episodes to run for each benchmark
seeds : list[int]
List of seeds to runs all benchmarks for. If None (default) seeds [1, ..., 10] are used.
"""
if bench is None:
bench = map(benchmarks.__dict__.get, benchmarks.__all__)
else:
bench = [getattr(benchmarks, b) for b in bench]
seeds = seeds if seeds is not None else range(10)
for b in bench:
print(f"Evaluating {b.__name__}")
for i in seeds:
print(f"Seed {i}/10")
bench = b()
try:
env = bench.get_benchmark(seed=i)
except:
continue
logger = Logger(
experiment_name=f"seed_{i}",
output_path=Path(results_path) / f"{b.__name__}",
)
perf_logger = logger.add_module(PerformanceTrackingWrapper)
logger.add_benchmark(bench)
logger.set_env(env)
env = PerformanceTrackingWrapper(env, logger=perf_logger)
agent = agent_method(env)
logger.add_agent(agent)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def run_random(results_path, benchmark_name, num_episodes, seeds, fixed):
bench = getattr(benchmarks, benchmark_name)()
for s in seeds:
if fixed > 1:
experiment_name = f"random_fixed{fixed}_{s}"
else:
experiment_name = f"random_{s}"
logger = Logger(experiment_name=experiment_name,
output_path=results_path / benchmark_name)
env = bench.get_benchmark(seed=s)
env = PerformanceTrackingWrapper(
env, logger=logger.add_module(PerformanceTrackingWrapper))
agent = DynamicRandomAgent(env, fixed)
logger.add_agent(agent)
logger.add_benchmark(bench)
logger.set_env(env)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def run_optimal(results_path,
benchmark_name,
num_episodes,
seeds=np.arange(10)):
bench = getattr(benchmarks, benchmark_name)()
if benchmark_name == "LubyBenchmark":
policy = optimal_luby
elif benchmark_name == "SigmoidBenchmark":
policy = optimal_sigmoid
elif benchmark_name == "FastDownwardBenchmark":
policy = optimal_fd
elif benchmark_name == "CMAESBenchmark":
policy = csa
else:
print("No comparison policy found for this benchmark")
return
for s in seeds:
if benchmark_name == "CMAESBenchmark":
experiment_name = f"csa_{s}"
else:
experiment_name = f"optimal_{s}"
logger = Logger(experiment_name=experiment_name,
output_path=results_path / benchmark_name)
env = bench.get_benchmark(seed=s)
env = PerformanceTrackingWrapper(
env, logger=logger.add_module(PerformanceTrackingWrapper))
agent = GenericAgent(env, policy)
logger.add_agent(agent)
logger.add_benchmark(bench)
logger.set_env(env)
logger.set_additional_info(seed=s)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def run_dacbench(results_path, agent_method, num_episodes):
"""
Run all benchmarks for 10 seeds for a given number of episodes with a given agent and save result
Parameters
-------
results_path : str
Path to where results should be saved
agent_method : function
Method that takes an env as input and returns an agent
num_episodes : int
Number of episodes to run for each benchmark
"""
for b in map(benchmarks.__dict__.get, benchmarks.__all__):
print(f"Evaluating {b.__name__}")
for i in range(10):
print(f"Seed {i}/10")
bench = b()
env = bench.get_benchmark(seed=i)
logger = Logger(
experiment_name=f"seed_{i}",
output_path=Path(results_path) / f"{b.__name__}",
)
perf_logger = logger.add_module(PerformanceTrackingWrapper)
logger.add_benchmark(bench)
logger.set_env(env)
logger.set_additional_info(seed=i)
env = PerformanceTrackingWrapper(env, logger=perf_logger)
agent = agent_method(env)
logger.add_agent(agent)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def run_static(results_path,
benchmark_name,
action,
num_episodes,
seeds=np.arange(10)):
bench = getattr(benchmarks, benchmark_name)()
for s in seeds:
logger = Logger(
experiment_name=f"static_{action}_{s}",
output_path=results_path / benchmark_name,
)
env = bench.get_benchmark(seed=s)
env = PerformanceTrackingWrapper(
env, logger=logger.add_module(PerformanceTrackingWrapper))
agent = StaticAgent(env, action)
logger.add_agent(agent)
logger.add_benchmark(bench)
logger.set_env(env)
logger.set_additional_info(seed=s, action=action)
run_benchmark(env, agent, num_episodes, logger)
logger.close()
def test_logging(self):
temp_dir = tempfile.TemporaryDirectory()
episodes = 5
logger = Logger(
output_path=Path(temp_dir.name),
experiment_name="test_logging",
)
bench = LubyBenchmark()
env = bench.get_environment()
time_logger = logger.add_module(EpisodeTimeWrapper)
wrapped = EpisodeTimeWrapper(env, logger=time_logger)
agent = StaticAgent(env=env, action=1)
run_benchmark(wrapped, agent, episodes, logger)
logger.close()
logs = load_logs(time_logger.get_logfile())
dataframe = log2dataframe(logs, wide=True)
# all steps must have logged time
self.assertTrue((~dataframe.step_duration.isna()).all())
# each episode has a recored time
episodes = dataframe.groupby("episode")
last_steps_per_episode = dataframe.iloc[episodes.step.idxmax()]
self.assertTrue(
(~last_steps_per_episode.episode_duration.isna()).all())
# episode time equals the sum of the steps in episode
calculated_episode_times = episodes.step_duration.sum()
recorded_episode_times = last_steps_per_episode.episode_duration
self.assertListEqual(calculated_episode_times.tolist(),
recorded_episode_times.tolist())
temp_dir.cleanup()
from pathlib import Path
from dacbench.agents import RandomAgent
from dacbench.logger import Logger
from dacbench.runner import run_benchmark
from dacbench.benchmarks import CMAESBenchmark
from dacbench.wrappers import StateTrackingWrapper
# Make CMAESBenchmark environment
bench = CMAESBenchmark()
env = bench.get_environment()
# Make Logger object to track state information
logger = Logger(experiment_name=type(bench).__name__,
output_path=Path("../plotting/data"))
logger.set_env(env)
# Wrap env with StateTrackingWrapper
env = StateTrackingWrapper(env, logger=logger.add_module(StateTrackingWrapper))
# Run random agent for 5 episodes and log state information to file
# You can plot these results with the plotting examples
agent = RandomAgent(env)
run_benchmark(env, agent, 5, logger=logger)
logger.close()
