def test_get_horizon():
env = examples.double_reentrant_line_model()
discount_factor = 0.9999
overrides = {
'BigStepPenaltyPolicyParams': {
'boolean_action_flag': True, # This should throw an error.
'convex_solver': 'cvx.CPLEX'
},
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepPolicy',
'horizon_mpc_ratio': 0,
'minimum_horizon': 1
}
}
ac_params, wk_params, si_params, po_params, hh_params, _, dp_params, _ \
= get_hedgehog_hyperparams(**overrides)
pf_mip_agent = PureFeedbackMIPHedgehogAgent(
env,
discount_factor,
wk_params,
hh_params,
ac_params,
si_params,
po_params,
demand_planning_params=dp_params)
assert pf_mip_agent.get_horizon() == 1
def test_get_hedgehog_hyperparams_overwrite_multiple():
overrides = {
'BigStepPenaltyPolicyParams': {
'boolean_action_flag': False,
'convex_solver': 'fake_solver',
'safety_penalty_coeff': 21,
'nonidling_penalty_coeff': 12
},
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepPenaltyPolicy',
'horizon_drain_time_ratio': 0.99,
'horizon_mpc_ratio': 0.55,
'minimum_horizon': 0.77,
'mpc_policy_class_name': 'fake_mpc_class',
'theta_0': 99
},
'name': 'funny_name'
}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
po_params = BigStepPenaltyPolicyParams(**overrides['BigStepPenaltyPolicyParams'])
hh_params = HedgehogHyperParams(**overrides['HedgehogHyperParams'])
name = overrides['name']
ac_params, wk_params, si_params, _, _, si_class, dp_params, _ \
= get_hedgehog_default_values('BigStepPenaltyPolicy')
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def test_get_hedgehog_hyperparams_overwrite_name():
overrides = {'name': "funny_name"}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
name = overrides['name']
ac_params, wk_params, si_params, po_params, hh_params, si_class, dp_params, _ \
= get_hedgehog_default_values('BigStepLayeredPolicy')
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def test_get_hedgehog_hyperparams_overwrite_asymptotic_cov():
overrides = {'AsymptoticCovarianceParams': {'num_presimulation_steps': 2, 'num_batch': 1}}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
ac_params = AsymptoticCovarianceParams(**overrides['AsymptoticCovarianceParams'])
_, wk_params, si_params, po_params, hh_params, si_class, dp_params, name \
= get_hedgehog_default_values()
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params,
po_params, hh_params, si_class, dp_params, name)
def test_get_hedgehog_hyperparams_overwrite_workload():
overrides = {'WorkloadRelaxationParams': {'num_vectors': 2, 'load_threshold': 0.1,
'convex_solver': 'fake_solver'}}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
wk_params = WorkloadRelaxationParams(**overrides['WorkloadRelaxationParams'])
ac_params, _, si_params, po_params, hh_params, si_class, dp_params, name \
= get_hedgehog_default_values()
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def test_initialised_si_object(self):
env = self.get_klimov_environment()
ac_params, wk_params, si_params, po_params, hh_params, _, dp_params, name \
= get_hedgehog_hyperparams()
agent = BigStepHedgehogGTOAgent(env, 1, wk_params, hh_params,
ac_params, po_params, dp_params, name)
agent.perform_offline_calculations()
assert isinstance(agent.strategic_idling_object, StrategicIdlingGTO)
def test_get_hedgehog_hyperparams_overwrite_demand_planning_params():
overrides = {
'DemandPlanningParams': {
'demand_planning_class_name': 'fake_class',
'params_dict': {'fake_key_1': 1, 'fake_key_2': 2}
}
}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
dp_params = DemandPlanningParams(**overrides['DemandPlanningParams'])
ac_params, wk_params, si_params, po_params, hh_params, si_class, _, name \
= get_hedgehog_default_values()
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def test_get_hedgehog_hyperparams_overwrite_big_step_layered_policy():
overrides = {
'BigStepLayeredPolicyParams': {
'convex_solver': 'fake_solver'
},
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepLayeredPolicy'
}
}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
po_params = BigStepLayeredPolicyParams(**overrides['BigStepLayeredPolicyParams'])
ac_params, wk_params, si_params, _, hh_params, si_class, dp_params, name \
= get_hedgehog_default_values('BigStepLayeredPolicy')
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def run_simulations(
num_sim: int, num_sim_steps: int, env: crw.ControlledRandomWalk,
discount_factor: float
) -> Tuple[List[snc_types.ActionSpace], List[snc_types.StateSpace]]:
""" Run multiple simulations on a given model and return all the actions and states.
:param num_sim: The number of simulations to run.
:param num_sim_steps: The number of simulation steps to run for each simulation.
:param env: the environment to stepped through.
:param discount_factor: discount factor used to compute the long term cost function.
"""
data_actions = [] # type: List[snc_types.ActionSpace]
data_states = [] # type: List[snc_types.StateSpace]
num_steps = num_sim * num_sim_steps
# Set Up Handlers
handlers = [
ProgressBarHandler(num_simulation_steps=num_steps, trigger_frequency=1)
] # type: List[Handler]
# Create Reporter
reporter = rep.Reporter(handlers=handlers) # fill with handlers
# Agent parameters
overrides: Dict[str, Dict[str, Union[str, float]]] = {}
ac_params, wk_params, si_params, po_params, hh_params, name, include_hedging \
= load_agents.get_hedgehog_hyperparams(**overrides)
for i in np.arange(num_sim):
job_gen_seed = int(42 + i)
np.random.seed(job_gen_seed)
# Create Policy Simulator
agent = BigStepHedgehogAgent(env, discount_factor, wk_params,
hh_params, ac_params, si_params,
po_params, include_hedging, name)
simulator = ps.SncSimulator(env,
agent,
discount_factor=discount_factor)
# Run Simulation
data = simulator.run(num_simulation_steps=num_sim_steps,
reporter=reporter,
job_gen_seed=job_gen_seed)
data_actions.extend(data["action"])
data_states.extend(data["state"])
return data_actions, data_states
def test_get_hedgehog_hyperparams_overwrite_strategic_idling():
overrides = {'StrategicIdlingParams': {
'strategic_idling_class': 'StrategicIdlingHedgehogNaiveGTO',
'convex_solver': 'fake_solver',
'epsilon': 5e6,
'shift_eps': 0.33,
'hedging_scaling_factor': 30,
'penalty_coeff_w_star': 3}}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
si_params = StrategicIdlingParams(**overrides['StrategicIdlingParams'])
si_class = load_agents.get_strategic_idling_class(
cast(StrategicIdlingParams, si_params).strategic_idling_class)
ac_params, wk_params, _, po_params, hh_params, _, dp_params, name \
= get_hedgehog_default_values()
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def build_default_simple_reentrant_line_simulator(seed):
"""
Helper function that returns a simulator to be used by the tests below.
"""
env = examples.simple_reentrant_line_model(job_gen_seed=seed)
overrides = {}
ac_params, wk_params, si_params, po_params, hh_params, si_class, dp_params, name \
= load_agents.get_hedgehog_hyperparams(**overrides)
# Create Policy Simulator
discount_factor = 0.95
agent = BigStepHedgehogAgent(env, discount_factor, wk_params, hh_params,
ac_params, si_params, po_params, si_class,
dp_params, name)
return ps.SncSimulator(env, agent, discount_factor=discount_factor)
def test_get_hedgehog_hyperparams_overwrite_hedgehog_hyperparams():
overrides = {
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepPenaltyPolicy',
'horizon_drain_time_ratio': 0.99,
'horizon_mpc_ratio': 0.55,
'minimum_horizon': 0.77,
'mpc_policy_class_name': 'fake_mpc_class',
'theta_0': 0.12
}
}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
hh_params = HedgehogHyperParams(**overrides['HedgehogHyperParams'])
ac_params, wk_params, si_params, po_params, _, si_class, dp_params, name \
= get_hedgehog_default_values('BigStepPenaltyPolicy')
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def construct_env_and_big_step_hedgehog_agent(override_dp_params):
env = examples.simple_reentrant_line_with_demand_model(job_gen_seed=42)
ac_params, wk_params, si_params, po_params, hh_params, si_class, dp_params, name \
= get_hedgehog_hyperparams(**override_dp_params)
agent = BigStepHedgehogAgent(
env,
0.9999,
wk_params,
hh_params,
ac_params,
si_params,
po_params,
si_class,
dp_params,
name
)
return env, agent
def test_get_hedgehog_hyperparams_overwrite_big_step_penalty_policy():
overrides = {
'BigStepPenaltyPolicyParams': {
'convex_solver': 'fake_solver',
'boolean_action_flag': True,
'safety_penalty_coeff': 33,
'nonidling_penalty_coeff': 55
},
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepPenaltyPolicy'
}
}
updated_params = load_agents.get_hedgehog_hyperparams(**overrides)
po_params = BigStepPenaltyPolicyParams(**overrides['BigStepPenaltyPolicyParams'])
ac_params, wk_params, si_params, _, hh_params, si_class, dp_params, name \
= get_hedgehog_default_values('BigStepPenaltyPolicy')
assert_equal_hedgehog_parameters(updated_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
def test_get_horizon():
env = examples.double_reentrant_line_model()
discount_factor = 0.9999
overrides = {
'HedgehogHyperParams': {
'horizon_mpc_ratio': 0,
'minimum_horizon': 1
}
}
ac_params, wk_params, si_params, _, hh_params, _, _, dp_params \
= get_hedgehog_hyperparams(**overrides)
pf_mip_agent = PureFeedbackStationaryHedgehogAgent(
env,
discount_factor,
wk_params,
hh_params,
ac_params,
si_params,
demand_planning_params=dp_params)
assert pf_mip_agent.get_horizon() == 1
def test_scenario(scenario_name, agent_class):
""" Run a brief integration test on a given scenario
"""
skip_tests = SKIPPED_TESTS + PULL_MODELS + PUSH_PULL_MODELS
if scenario_name in skip_tests:
pytest.skip()
np.random.seed(SEED_NO)
_, env = scenarios.load_scenario(scenario_name, SEED_NO)
# Update parameters for quick tests.
overrides = {
"HedgehogHyperParams": {
"theta_0": 0.5,
"horizon_drain_time_ratio": 0.1,
"horizon_mpc_ratio": 0.1,
"minimum_horizon": 10
},
"AsymptoticCovarianceParams": {
"num_presimulation_steps": 100,
"num_batch": 20
}
}
if scenario_name in MIP_REQUIRED_MODELS:
overrides["HedgehogHyperParams"]["mpc_policy_class_name"] = "FeedbackMipFeasibleMpcPolicy"
ac_params, wk_params, si_params, po_params, hh_params, si_class, dp_params, name \
= load_agents.get_hedgehog_hyperparams(**overrides)
discount_factor = 0.95
if agent_class == BigStepHedgehogAgent:
agent = agent_class(env, discount_factor, wk_params, hh_params, ac_params,
si_params, po_params, si_class, dp_params, name)
elif agent_class == BigStepHedgehogGTOAgent:
agent = agent_class(env, discount_factor, wk_params, hh_params, ac_params,
po_params, dp_params, name)
else:
assert False, f"Not recognised agent: {agent_class}"
simulator = ps.SncSimulator(env, agent, discount_factor=discount_factor)
simulator.run(num_simulation_steps=SIM_STEPS)
def test_create_agent_with_mip_flag_set_to_false():
env = examples.double_reentrant_line_model()
discount_factor = 0.9999
overrides = {
'BigStepPenaltyPolicyParams': {
'boolean_action_flag': False, # This should throw an error.
'convex_solver': 'cvx.CPLEX'
},
'HedgehogHyperParams': {
'activity_rates_policy_class_name': 'BigStepPolicy'
}
}
ac_params, wk_params, si_params, po_params, hh_params, _, dp_params, _ \
= get_hedgehog_hyperparams(**overrides)
with pytest.raises(AssertionError):
_ = PureFeedbackMIPHedgehogAgent(env,
discount_factor,
wk_params,
hh_params,
ac_params,
si_params,
po_params,
demand_planning_params=dp_params)
def test_get_hedgehog_hyperparams_default_values():
default_params = load_agents.get_hedgehog_hyperparams()
ac_params, wk_params, si_params, po_params, hh_params, si_class, dp_params, name \
= get_hedgehog_default_values()
assert_equal_hedgehog_parameters(default_params, ac_params, wk_params, si_params, po_params,
hh_params, si_class, dp_params, name)
