Motivation

The current GPSampler does not support constant liar and hence is not very effective for batched / asynchronous hyperparam suggestions.

Description

Maybe also extend the constant_liar feature to the GPSampler?

Alternatives (optional)

No response

Additional context (optional)

In the book 'Bayesian Optimization' by Roman Garnett, page 255. The sequential simulation + constant liar strategy is mentioned, e.g. imputing value with avg(y_observed) and claimed to have good performance.