Non-myopic Sequential Bayesian Experimental Design

This direction is merged into another direction: myopic acquisition functions. Its last version stays readable and citable here.

Version 1 of 191 papers110 conceptsfrozen 2026-10-04

In scientific discovery, the most expensive resource is often the experiment itself. Whether you are synthesizing new materials, testing drug efficacy, or probing the interior of a volcano, you want every data point to count. The traditional framework for this is Bayesian Optimal Experimental Design (BOED), which uses a "greedy" strategy: it selects the next experiment that maximizes the immediate Expected Information Gain (EIG).

However, a growing body of research argues that this greedy approach is fundamentally limited. By focusing only on the immediate next step, researchers fall into a "myopia trap," missing out on complex, multi-step strategies that yield far more information over the long run. The shift toward non-myopic sequential design treats experimental discovery not as a series of isolated choices, but as a long-horizon planning problem.

The Myopia Trap

Standard BOED is "myopic" because it assumes the best way to learn about a system is to maximize information now. As noted by Adam Foster and Tom Rainforth in Deep Adaptive Design, this shortsightedness prevents the discovery of strategies where a "boring" or low-information experiment today unlocks a massive discovery tomorrow.

Consider a robot navigating a physical space to map a phenomenon. A myopic agent might move to the nearest point of high uncertainty. A non-myopic agent, however, might realize that traveling further now—even if it collects less data along the way—positions it to sweep a much more informative region later. Xun Huan and Youssef Marzouk demonstrated in Sequential Bayesian optimal experimental design via approximate dynamic programming that such foresight can double the expected information gain compared to greedy methods.

Framing Design as Sequential Planning

To move beyond myopia, researchers have reframed experimental design as a Markov Decision Process (MDP) or a Partially Observable MDP (POMDP). In this view, the "state" is our current belief about the world (the posterior distribution), and the "action" is the choice of the next experiment.

This framing allows the use of powerful tools from reinforcement learning and control theory:

  • Dynamic Programming: Xun Huan and Youssef Marzouk applied Approximate Dynamic Programming (ADP) to backward-induct the value of future experiments.
  • Look-ahead Search: Shali Jiang introduced One-Shot Multi-Step Trees, which collapse a multi-step decision tree into a single optimization problem, making long-term planning computationally feasible.
  • Monte Carlo Tree Search (MCTS): For complex, discrete spaces like malware analysis or cyber-forensics, researchers like Alessandro Farinelli and Soodeh Atefi use MCTS to simulate future interaction sequences, identifying paths that reveal hidden states.

Amortizing the Cost of Foresight

The primary barrier to non-myopic design is computational. Calculating the EIG for a single step is hard; calculating it for a sequence of ten steps is exponentially harder.

Deep Adaptive Design (DAD), pioneered by Adam Foster and colleagues, solves this by "amortizing" the cost. Instead of solving a complex optimization problem during the experiment, DAD trains a neural network policy offline. This policy learns to map current experimental results directly to the next optimal design. Because the network is trained to maximize total information gain over a full trajectory, it naturally develops non-myopic behaviors—such as broad initial exploration followed by targeted refinement—that can be executed in milliseconds during a live trial.

Scaling to High Dimensions

While early non-myopic methods worked only for simple models, recent work has scaled these ideas to high-dimensional spaces like image reconstruction and biological sequence design.

Jacopo Iollo and Florence Forbes introduced CoDiff, which uses diffusion models to handle high-dimensional priors. Similarly, DiffBED uses diffusion to regularize the design process, preventing the optimizer from "reward hacking" and ensuring that the planned experiments remain physically realistic and informative across multiple steps.

Beyond Simple Information

The non-myopic perspective is also transforming specialized fields:

  • Causal Discovery: Yuxuan Wang showed in Bayesian Active Learning for Bivariate Causal Discovery that multi-step look-ahead is essential for causal inference. It allows a system to balance immediate evidence gathering with the long-term goal of distinguishing cause from effect, requiring fewer expensive interventions.
  • Robustness: Standard BOED often fails if the model is slightly wrong. Yaman Barlas and Samuel Kaski have proposed Generalised BOED, which uses robust loss functions to ensure that sequential plans don't collapse when faced with outliers or model misspecification.

By treating experimental design as a strategic, long-term endeavor rather than a series of greedy choices, this research direction is enabling more efficient discovery in everything from robotics to biological sequence design. The reader should take away that the "best" next experiment is rarely the one that looks best in isolation; it is the one that best serves the entire remaining experimental budget.

Papers in this version

Showing 60 of 91 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_myopic_sequential_bayesian_experimen_0c2504cf,
  title = {Non-myopic Sequential Bayesian Experimental Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-myopic-sequential-bayesian-experimental-design/rdr_0c2504cf01484499967d505aef32eab0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_myopic_sequential_bayesian_experimen_0c2504cf_v1,
  title = {Non-myopic Sequential Bayesian Experimental Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-myopic-sequential-bayesian-experimental-design/rdr_0c2504cf01484499967d505aef32eab0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}