I am a research scientist at Meta on the Core Data Science team, PhD candidate in machine learning at the Universtiy of Oxford, and co-creator of BoTorch---an open source library for Bayesian optimization research. Within Core Data Science, I work in the Adaptive Experimentation research group. I am a member of the Machine Learning Research Group at Oxford. During my PhD, I am working with Michael Osborne (Oxford), Eytan Bakshy (Meta), and Max Balandat (Meta). My research focuses on methods for principled, sample-efficient optimization including Bayesian optimization and transfer learning. I am particularly interested in practical methods for principled exploration (using probablistic models) that are are robust across applied problems and depend on few, if any, hyperparameters. Furthermore, I aim to democratize such methods by open sourcing reproducible code. Prior to joining Meta, I worked with Finale Doshi-Velez at Harvard University on efficient and robust methods for transfer learning.

In my free time, it's a safe bet that I'm climbing, skiing, running, scuba diving, or scheming about how to get to the remote reaches of the world.

Selected Publications


Conference Publications

Samuel (Sam) Daulton

Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization
We propose a new state-of-the-art acquisition function for constrained Bayesian optimization of multiple competing objectives. The acquisition function is an exact calculation of expected hypervolume improvement other than monte-carlo integration error. It enables parallel (batch) evaluations via the inclusion-exclusion principle, and it is differentiable, which enables efficient gradient-based acquisition optimization. We prove that the sample average gradient of the Monte-Carlo acquisition is an unbiased estimator of the gradient of the true expected hypervolume improvement.
Samuel Daulton, Maxmilian Balandat, Eytan Bakshy
Advances in Neural Information Processing Systems 33, 2020.
Paper | Code | Video

Samuel (Sam) Daulton

BoTorch: Programmable Bayesian Optimization in PyTorch
We propose a modular Monte-Carlo-based framework for developing new methods for Bayesian optimization. We include multiple examples including a novel one-shot optimization formulation of the Knowledge Gradient. We provide convergence guarantees for a broad class of quasi-Monte-Carlo acquisition functions using the sample average approximation.
Maxmilian Balandat, Brian Karrer, Daniel Jiang, Samuel Daulton, Benjamin Letham, Andrew Gordon Wilson, Eytan Bakshy
Advances in Neural Information Processing Systems 33, 2020.
Paper | Code


Samuel (Sam) Daulton

Robust and Efficient Transfer Learning with Hidden Parameter Markov Decision Processes
We propose a novel algorithm for transfering knowledge between similar, but different reinforcement learning tasks by learning low-dimensional latent embeddings. These embeddings encode task-specific nuances and are provided with the state and action as the input to transition model that is shared across task instances. Modeling the dynamics with a Bayesian neural network enables scalable inference and joint optimization of the network parameters and latent embeddings.
Taylor W. Killian*, Samuel Daulton* (*equal contribution), George Konidaris, Finale Doshi-Velez
[Oral] Advances in Neural Information Processing Systems 30, 2017.
Paper | Code

Workshop Papers

Samuel (Sam) Daulton

Distilled Thompson Sampling: Practical and Efficient Thompson Sampling via Imitation Learning
We propose a practical method for distilling a Thompson sampling policy into a compact, explicit policy representation for contextual bandit optimization in applications with limited memory and low-latency requirements. The expensive posterior sampling and numerical optimization is performed offline, while the imitation policy is used for efficient online decision-making. We show that our method enjoys the same Bayes regret as the best UCB algorithm, up to a sum of single time step approximation errors, which can be easily controlled with abudant "unlabeled" contexts (without an action or reward) that are available in many practical applications (e.g. firms typically have databases with features about different entities).
Hongseok Namkoong*, Samuel Daulton* (*equal contribution), Eytan Bakshy
[Oral] NeurIPS Offline Reinforcement Learning Workshop, 2020.
Paper | Video | Talk


Samuel (Sam) Daulton

Thompson Sampling for Contextual Bandit Problems with Auxiliary Safety Constraints
We consider a new class of contextual bandit problems with constraints on auxiliary outcome(s). We consider upper confidence bound and Thompson sampling-based algorithms and perform ablation studies revealing nice properties regarding fairness using the Thompson sampling algorithm. We demonstrate the performance of the algorithm on a real world video transcoding problem.
Samuel Daulton, Shaun Singh, Vashist Avadhanula, Drew Dimmery, Eytan Bakshy
NeurIPS Workshop on Safety and Robustness in Decision Making, 2019.
Paper




Talks


Samuel (Sam) Daulton

Practical Solutions to Exploration Problems
I discuss many practical approaches we use at Facebook for principled exploration including policy optimization with Bayesian optimization via online experiments, constrained Bayesian optimization, multi-task Bayesian optimization accelerated with offline simulations, and contextual bandits.
Samuel Daulton
Netflix ML Platform Meetup on Exploration and Exploitation, 2019.
Video | Slides