Query: Potential Negative Values for kappa and xi Due to exploration_decay

[spline2hg]

In the UpperConfidenceBound, ProbabilityOfImprovement, and ExpectedImprovement classes, the parameter kappa is expected to be greater than or equal to 0, and xi should be positive(though no checks for this). While there is a check ensuring that kappa is initialized with a non-negative value, there seems to be no such validation for the exploration_decay parameter.

If exploration_decay is set to a negative value, it could potentially cause kappa or xi to become negative over time. Is this an intended or desirable behavior? Should there be an additional check or constraint to prevent this from happening?

[till-m]

Hi @spline2hg,

this isn't exactly about desirable behaviour so much as there being a finite limit to the amount of guardrails we can implement. No matter how you program your tool, there will probably always be someone shooting themselves in the foot with it, provided your userbase is big enough.

That being said, if you wish to make a PR and add something to this effect, feel free to tag me and I will review it :)

[spline2hg]

Hi @till-m,

Would it be appropriate to use if/else checks in the init, similar to how we handle kappa validation in Upper Confidence Bound, or is there a better approach for handling this type of validation?

I’d be happy to work on a Pull Request for this.

[till-m]

Hey @spline2hg,

I think that's a fine place (also, FYI, the value should probably be 0 < decay <= 1) . If you're at it, maybe you can also add a check for xi?

Thanks!

[till-m]

closed in #577 -- thanks!