Apply Prior Knowledge
You can leverage SigOpt to take advantage of your domain expertise on how metric values behave for certain parameters. These prior beliefs make the optimization process more efficient.
Defining the Prior Distribution
Knowledge of Prior Distributions can be defined through the
prior field for each continuous parameter. By default, SigOpt assumes that all parameters have a uniform prior distribution. You can inspect the probability density function of the prior belief distributions and generate the corresponding code snippet using the interactive tool below.When the prior belief is set for a parameter, SigOpt is more likely to generate configurations from regions of this parameter with high probability density (PDF) value earlier in the SigOpt Experiment. Generally speaking, parameter configurations with PDF value
2 are twice as likely to be suggested as those with PDF value 1. The effect of prior belief is most notable during the initial portion of an experiment.Example: Normally Distributed Parameters
Suppose in prior experiments, you observe that the
log_learning_rate parameter exhibits properties indicating that the highest performing values are normally distributed.
Defining Prior Distributions with Code
Normal Distribution:
1
dict(
2
name="log_learning_rate",
3
bounds=dict(min=np.log(0.000001), max=np.log(1)),
4
prior=dict(name="normal", mean=-1.5, scale=1),
5
type="double",
6
)
Copied!
Beta Distribution:
1
dict(
2
name="beta_distribution",
3
bounds=dict(min=0, max=1),
4
prior=dict(name="beta", shape_a=2, shape_b=4.5),
5
type="double",
6
)
Copied!
Limitations
Last modified 24d ago