Optimize new plumbing (#101)

* updating to new interventions and allowing fixed

* minor update to test case

* minor update to testcase

* adding initial_guess to test

* adding missing objective_function_option

* correcting converting indexing

service/models/base.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import ClassVar, Dict
+from typing import ClassVar, Dict, Optional
 from pydantic import BaseModel, Field
 
 
@@ -42,6 +42,31 @@ class InterventionSelection(BaseModel):
  name: str
 
 
+class HMIStaticIntervention(BaseModel):
+ timestep: float
+ value: float
+
+
+class HMIDynamicIntervention(BaseModel):
+ parameter: str
+ threshold: float
+ value: float
+ is_greater_than: bool
+
+
+class HMIInterventionPolicy(BaseModel):
+ model_id: Optional[str] = Field(default="")
+ Interventions: list[HMIIntervention]
+
+
+class HMIIntervention(BaseModel):
+ name: str
+ applied_to: str
+ type: str
+ static_interventions: Optional[list[HMIStaticIntervention]] = Field(default=None)
+ dynamic_interventions: Optional[list[HMIDynamicIntervention]] = Field(default=None)
+
+
 class OperationRequest(BaseModel):
  pyciemss_lib_function: ClassVar[str] = ""
  engine: str = Field("ciemss", example="ciemss")

service/models/converters.py

@@ -4,14 +4,23 @@
 import torch
 from utils.tds import fetch_interventions
 from typing import Dict
+from models.base import HMIIntervention
 
 
 def fetch_and_convert_static_interventions(policy_intervention_id, job_id):
- static_interventions: Dict[torch.Tensor, Dict[str, any]] = defaultdict(dict)
  if not (policy_intervention_id):
- return static_interventions
+ return defaultdict(dict)
  policy_intervention = fetch_interventions(policy_intervention_id, job_id)
- for inter in policy_intervention["interventions"]:
+ interventionList = policy_intervention["interventions"]
+ return convert_static_interventions(interventionList)
+
+
+# Used to convert from HMI Intervention Policy -> pyciemss static interventions.
+def convert_static_interventions(interventions: list[HMIIntervention]):
+ if not (interventions):
+ return defaultdict(dict)
+ static_interventions: Dict[torch.Tensor, Dict[str, any]] = defaultdict(dict)
+ for inter in interventions:
  for static_inter in inter["static_interventions"]:
  time = torch.tensor(float(static_inter["timestep"]))
  parameter_name = inter["applied_to"]

service/models/operations/optimize.py

@@ -6,14 +6,14 @@
 import numpy as np
 import torch
 from pydantic import BaseModel, Field, Extra
-from models.base import OperationRequest, Timespan
+from models.base import OperationRequest, Timespan, HMIIntervention
 from pyciemss.integration_utils.intervention_builder import (
  param_value_objective,
  start_time_objective,
 )
 
 from pyciemss.ouu.qoi import obs_nday_average_qoi, obs_max_qoi
-from models.converters import fetch_and_convert_static_interventions
+from models.converters import convert_static_interventions
 from utils.tds import fetch_model, fetch_inferred_parameters
 
 
@@ -37,22 +37,26 @@ def gen_call(self):
  return qoi_map[self.method]
 
 
-def objfun(x, is_minimized):
- if is_minimized:
+def objfun(x, initial_guess, objective_function_option):
+ if objective_function_option == "lower_bound":
  return np.sum(np.abs(x))
- else:
+ if objective_function_option == "upper_bound":
  return -np.sum(np.abs(x))
+ if objective_function_option == "initial_guess":
+ return np.sum(np.abs(x - initial_guess))
 
 
 class InterventionObjective(BaseModel):
- selection: str = Field(
+ intervention_type: str = Field(
  "param_value",
  description="The intervention objective to use",
  example="param_value",
  )
  param_names: list[str]
  param_values: Optional[list[Optional[float]]] = None
  start_time: Optional[list[float]] = None
+ objective_function_option: Optional[list[str]] = None
+ initial_guess: Optional[list[float]] = None
 
 
 class OptimizeExtra(BaseModel):
@@ -70,7 +74,6 @@ class OptimizeExtra(BaseModel):
  )
  maxiter: int = 5
  maxfeval: int = 25
- is_minimized: bool = True
  alpha: float = 0.95
  solver_method: str = "dopri5"
  solver_options: Dict[str, Any] = {}
@@ -80,7 +83,10 @@ class Optimize(OperationRequest):
  pyciemss_lib_function: ClassVar[str] = "optimize"
  model_config_id: str = Field(..., example="ba8da8d4-047d-11ee-be56")
  timespan: Timespan = Timespan(start=0, end=90)
- policy_interventions: InterventionObjective
+ optimize_interventions: InterventionObjective # These are the interventions to be optimized.
+ fixed_static_parameter_interventions: list[HMIIntervention] = Field(
+ None
+ ) # Theses are interventions provided that will not be optimized
  step_size: float = 1.0
  qoi: QOI
  risk_bound: float
@@ -90,35 +96,35 @@ class Optimize(OperationRequest):
  None,
  description="optional extra system specific arguments for advanced use cases",
  )
- policy_intervention_id: str = Field(None, example="ba8da8d4-047d-11ee-be56")
 
  def gen_pyciemss_args(self, job_id):
  # Get model from TDS
  amr_path = fetch_model(self.model_config_id, job_id)
- static_interventions = fetch_and_convert_static_interventions(
- self.policy_intervention_id, job_id
+ fixed_static_parameter_interventions = convert_static_interventions(
+ self.fixed_static_parameter_interventions
  )
 
- intervention_type = self.policy_interventions.selection
+ intervention_type = self.optimize_interventions.intervention_type
  if intervention_type == "param_value":
- assert self.policy_interventions.start_time is not None
+ assert self.optimize_interventions.start_time is not None
  start_time = [
- torch.tensor(time) for time in self.policy_interventions.start_time
+ torch.tensor(time) for time in self.optimize_interventions.start_time
  ]
- param_value = [None] * len(self.policy_interventions.param_names)
+ param_value = [None] * len(self.optimize_interventions.param_names)
 
- policy_interventions = param_value_objective(
+ optimize_interventions = param_value_objective(
  start_time=start_time,
- param_name=self.policy_interventions.param_names,
+ param_name=self.optimize_interventions.param_names,
  param_value=param_value,
  )
  else:
- assert self.policy_interventions.param_values is not None
+ assert self.optimize_interventions.param_values is not None
  param_value = [
- torch.tensor(value) for value in self.policy_interventions.param_values
+ torch.tensor(value)
+ for value in self.optimize_interventions.param_values
  ]
- policy_interventions = start_time_objective(
- param_name=self.policy_interventions.param_names,
+ optimize_interventions = start_time_objective(
+ param_name=self.optimize_interventions.param_names,
  param_value=param_value,
  )
 
@@ -127,20 +133,23 @@ def gen_pyciemss_args(self, job_id):
  extra_options.pop("inferred_parameters"), job_id
  )
  n_samples_ouu = extra_options.pop("num_samples")
- is_minimized = extra_options.pop("is_minimized")
 
  return {
  "model_path_or_json": amr_path,
  "logging_step_size": self.step_size,
  "start_time": self.timespan.start,
  "end_time": self.timespan.end,
- "objfun": lambda x: objfun(x, is_minimized),
+ "objfun": lambda x: objfun(
+ x,
+ self.optimize_interventions.initial_guess[0],
+ self.optimize_interventions.objective_function_option[0],
+ ),
  "qoi": self.qoi.gen_call(),
  "risk_bound": self.risk_bound,
  "initial_guess_interventions": self.initial_guess_interventions,
  "bounds_interventions": self.bounds_interventions,
- "static_parameter_interventions": policy_interventions,
- "fixed_static_parameter_interventions": static_interventions,
+ "static_parameter_interventions": optimize_interventions,
+ "fixed_static_parameter_interventions": fixed_static_parameter_interventions,
  "inferred_parameters": inferred_parameters,
  "n_samples_ouu": n_samples_ouu,
  **extra_options,

tests/examples/optimize/input/request.json

@@ -2,11 +2,13 @@
  "engine": "ciemss",
  "user_id": "not_provided",
  "model_config_id": "sidarthe",
- "policy_interventions": {
- "selection": "param_value",
+ "optimize_interventions": {
+ "intervention_type": "param_value",
+ "objective_function_option": ["lower_bound"],
  "start_time": [2],
  "param_names": ["beta"],
- "param_values": [0.02]
+ "param_values": [0.02],
+ "initial_guess": [0]
  },
  "timespan": {
  "start": 0,