search

secondmind-labs / trieste

A Bayesian optimization toolbox built on TensorFlow
Apache License 2.0
224 stars 42 forks source link

fix for multi objective integration test #678

Closed hstojic closed 1 year ago

hstojic commented 1 year ago

we were computing ref_point on observations instead of the true pareto front

hstojic commented 1 year ago

actually in get_reference_point we extract the nadir:

    front = Pareto(observations).front
    f = tf.math.reduce_max(front, axis=-2) - tf.math.reduce_min(front, axis=-2)
    return tf.math.reduce_max(front, axis=-2) + 2 * f / tf.cast(tf.shape(front)[-2], f.dtype)

so I think this fix is not needed?

That nadir is based on acquired data, not on the true front - for evaluating the performance we should have a fixed one based on the true pareto front, right?