ef0d17b3bdb4ee2aa741ba28c7255c53-AuthorFeedback.pdf

PAML11 learns a representation of the space of tasks and makes comparisons directly in that space. This way our approach12 does not require a) rollouts on new potential tasks, b) handpicked reference tasks and c) the task parameters to be13 observeddirectly. In4.4,itusually24 starts to select lengths at both ends of the range and then selects tasks towards the middle.