In a successive querying problem, one achieves a certain objective by issuing a series of queries to an oracle and receives a series of observations in return. It is a challenging task because the queries need to be adaptive to the outcomes observed so-far, while being informative to the objective at hand. While successful algorithms have been developed for a range of successive querying tasks, these algorithms can be slow to compute and in some cases, intractable. A common Achilles's heel of these prior works is their reliance on the computation over the space of oracle functions itself during inference time. As a result, when the space of oracle functions becomes complex, these approaches become computationally infeasible.
In this thesis, we explore an alternative approach to informative query selection by performing meta-learning on related observation data: Given a set of observation data generated by similar oracle functions, we train a propagator in meta-learning time. A propagator is a function that maps outcomes from past observations to outcomes of future queries directly. We show that by simply propagating past observations to future outcomes, one can perform a range of successive querying applications that are previously intractable. To this end, we prescribe a general work-flow of informative querying with a propagator: In meta-learning time, a domain agnostic training process is used to train the propagator, and at inference time, a domain specific acquisition function is chosen to specialise the propagator in making informative queries in the specific domain of interest.
Thesis Supervisor(s): Prof. Armando Solar-Lezama and Leslie Kaelbling