def generate_nn_polyhedral_guard(self, nn, chosen_action, output_flag):
gurobi_model = grb.Model()
gurobi_model.setParam('OutputFlag', output_flag)
gurobi_model.setParam('Threads', 2)
observation = gurobi_model.addMVar(shape=(2, ),
lb=float("-inf"),
ub=float("inf"),
name="observation")
Experiment.generate_nn_guard(gurobi_model,
observation,
nn,
action_ego=chosen_action)
observable_template = Experiment.octagon(2)
self.env_input_size = 2
observable_result = self.optimise(observable_template, gurobi_model,
observation)
self.env_input_size = 6
return observable_template, observable_result
def run_parameterised_experiment(config):
# Hyperparameters
trial_dir = tune.get_trial_dir()
problem, method, other_config = config["main_params"]
n_workers = config["n_workers"]
experiment = CartpoleExperiment()
experiment.nn_path = other_config[
"folder"] # nn_paths_cartpole[other_config["nn_path"]]
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(
experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(
experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
# experiment.use_rounding = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
safe_value = 0
if safe is None:
safe_value = 0
elif safe:
safe_value = 1
elif not safe:
safe_value = -1
tune.report(elapsed_seconds=elapsed_seconds,
safe=safe_value,
max_t=max_t,
done=True)
def run_parameterised_experiment(config):
# Hyperparameters
trial_dir = tune.get_trial_dir()
problem, method, other_config = config["main_params"]
n_workers = config["n_workers"]
if problem == "bouncing_ball":
if method == "ora":
experiment = ORABouncingBallExperiment()
else:
experiment = BouncingBallExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_bouncing_ball[other_config["nn_path"]])
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
raise NotImplementedError()
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
elif problem == "stopping_car":
if method == "ora":
experiment = ORAStoppingCarExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_stopping_car[other_config["nn_path"]])
experiment.input_epsilon = other_config["epsilon_input"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
else:
experiment = StoppingCarExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_stopping_car[other_config["nn_path"]])
experiment.input_epsilon = other_config["epsilon_input"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
else:
if method == "ora":
experiment = ORACartpoleExperiment()
else:
experiment = CartpoleExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_cartpole[other_config["nn_path"]])
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
# experiment.use_rounding = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
safe_value = 0
if safe is None:
safe_value = 0
elif safe:
safe_value = 1
elif not safe:
safe_value = -1
tune.report(elapsed_seconds=elapsed_seconds, safe=safe_value, max_t=max_t, done=True)
def get_nn(self):
config = get_PPO_config(1234)
trainer = ppo.PPOTrainer(config=config)
trainer.restore(self.nn_path)
policy = trainer.get_policy()
sequential_nn = convert_ray_policy_to_sequential(policy).cpu()
# l0 = torch.nn.Linear(6, 2, bias=False)
# l0.weight = torch.nn.Parameter(torch.tensor([[0, 0, 1, -1, 0, 0], [1, -1, 0, 0, 0, 0]], dtype=torch.float32))
# layers = [l0]
# for l in sequential_nn:
# layers.append(l)
#
# nn = torch.nn.Sequential(*layers)
nn = sequential_nn
# ray.shutdown()
return nn
if __name__ == '__main__':
ray.init(log_to_driver=False, local_mode=False)
experiment = StoppingCarExperiment()
experiment.plotting_time_interval = 60 * 2
experiment.show_progressbar = True
experiment.show_progress_plot = False
template = Experiment.octagon(experiment.env_input_size)
experiment.analysis_template = template # standard
input_boundaries = [40, -30, 10, -0, 28, -28, 36, -36, 0, -0, 0, -0, 0]
experiment.input_boundaries = input_boundaries
experiment.time_horizon = 150
experiment.run_experiment()