Online Experiment Design

Fully Adaptive Optimization

The information gathered by updates of a posterior distribution, such as the updates generated by the SMC approximation, can be used to inform the choice of future experiments. The ExperimentDesigner encapsulates and automates this process by designing experiments that minimize an objective function

\[O(\vec{e}) = r(\vec{e}) + k C(\vec{e}),\]

where \(r(\vec{e})\) is the Bayes risk (see [GFWC12]), \(C(\vec{e})\) is a cost function describing the cost associated with each experiment and where \(k\) is a parameter describing how much we are willing to pay for reductions in risk. In QInfer, the cost function is specified as the experiment_cost() method of the Simulatable class under study.


As opposed to fully optimizing the utility of an experiment, it can be substantially less expensive to use a heuristic function of prior information to select experiments without explicit simulation. As an example, in [WGFC13a], the particle guess heuristic (PGH) was used to design quantum Hamiltonian learning experiments without incurring additional simulation costs.

QInfer exposes these heuristics though subclasses of Heuristic, such as PGH, which create experiment parameter arrays given the prior information exposed by an SMCUpdater instance.