Desi Ivanova

Deep Adaptive Design: Model-Based Adaptive Experimental Design in Real Time

Traditional model-based approaches for adaptive experimental design are typically too computationally costly to run in real time, making them unsuitable for applications that require quick design decisions. In this talk, we introduce Deep Adaptive Design (DAD), a general framework that reduces the cost of performing Bayesian adaptive experiments. Rather than optimizing designs directly during the experiment, DAD learns a design policy network upfront that is then used to perform multiple adaptive experiments at deployment time. The policy network takes past design-outcome pairs as an input and outputs the design for the next experiment iteration, enabling quick and adaptive design decisions with a single forward pass through the network. Remarkably, we find that DAD not only speeds up the adaptive experimentation process but can also significantly improve its performance by learning non-greedy strategies and avoiding errors resulting from inexact inference in the traditional framework.