Controlling the Conversation
Abstract
We study how AI assistants design their conversations. An assistant, with a private objective, guides a user through her search via sequential queries, modeled as partitions of the search space. The user selects the subset with the highest average value. Using a prior-free framework, we minimize the platform's regret relative to an omniscient benchmark.
We provide an optimal randomized policy that achieves bounded regret of $\frac{1}{n-1}$, where $n$ is the number of items. This policy is implemented as a sequence of ``recommendations'', that is, the AI sequentially proposes candidates from the item set until one is accepted.
We provide an optimal randomized policy that achieves bounded regret of $\frac{1}{n-1}$, where $n$ is the number of items. This policy is implemented as a sequence of ``recommendations'', that is, the AI sequentially proposes candidates from the item set until one is accepted.
Research Areas
