Skip to main contentIBM ST4SD

Using runtime policies

Use this page to learn how to use the ST4SD runtime policies, for example to execute parameterised packages that contain surrogate models.

What is a runtime policy?

Runtime policies are algorithms which modify the default behaviour of ST4SD when its asked to run a parameterised package.

How to use a runtime policy?

You can use a runtime policy by configuring the runtimePolicy field of the api_experiment_start() payload.

Example

from __future__ import typing
import json
def experiment_start(api: experiment.service.db.ExperimentRestAPI) -> str:
    """Start an experiment that uses the prefer-surrogate runtime policy


    Arguments:
      api: An instance of ExperimentRestAPI that has connected to your
           private ST4SD registry

Policy: prefer-surrogate

The algorithm of the policy is:

  1. Find a suitable surrogate package of the requested parameterised package in the api_experiment_start() payload
    1. Find parameterised packages that are the result of synthesizing a transformation relationship, AND
    2. The outputGraph of the associated transformation relationship matches a subGraph of the requested parameterised package.
  2. If there is a suitable surrogate parameterised package then start it and return the UID of the resulting run.
  3. Otherwise, start and return the UID of the requested parameterised packages.

The schema of the runtimePolicy field in the api_experiment_start() payload (in YAML format) is:

runtimePolicy:
  name: prefer-surrogate

Policy: random-canary-surrogate

In a future version the runtime system will automatically invoke this runtime-policy. However, users can also switch it on manually for their virtual experiment runs.

The algorithm of the policy is:

  1. Toss a weighted coin which shows heads with probability probabilitySurrogate. If the coin shows heads when it lands then attempt to start 2 virtual experiment instances. The requested one AND a surrogate of the requested one (if available).
    1. If coin shows heads: then attempt to start 2 virtual experiment instances. An instance using the requested parameterised package from the api_experiment_start() payload and a second one of a suitable surrogate parameterised package (if present). If there is no suitable surrogate parameterised package then start just one run of the requested parameterised package.
    2. Use the same logic as the “prefer-surrogate” to identify one surrogate parameterised package.
  2. If the coin shows tails then start just the requested parameterised package
  3. Return the UID of the instance for the parameterised package.

The schema of the runtimePolicy field in the api_experiment_start() payload (in YAML format) is:

runtimePolicy:
  name: random-canary-surrogate
  config: # optional
    probabilitySurrogate: float from 0.0 to 1.0 (optional, default is 0.5)
Edit this page on GitHub
Previous
ST4SD Services: Using Graph Relationships