Choosing an agent

banditry ships two agent families, each with two variants. All of them are built through build_agent(config, space).

OFU (GP-UCB) — OFUGPConfig

Optimism-in-the-face-of-uncertainty with a Gaussian-process surrogate: fit a GP, optimise a lower-confidence-bound acquisition.

Variant	When
surrogate="gp" (exact GP)	Small budgets (up to a few hundred observations). Exact inference, best sample-efficiency.
surrogate="svgp" (sparse variational GP)	Larger budgets where exact GP inference gets slow. Minibatch ELBO training with inducing points.

noise_std_proxy (the assumed observation-noise scale) is required. Setting frequentist=True switches the confidence width to the Chowdhury–Gopalan β_t, which additionally needs rkhs_norm.

from banditry import OFUGPConfig, build_agent
agent = build_agent(OFUGPConfig(surrogate="gp", noise_std_proxy=1.0), space)

Thompson sampling — TSConfig

A neural value function whose weights are drawn from the posterior by an MCMC oracle; each round acts greedily w.r.t. one posterior sample.

Variant	When
sampler="langevin" (SGLD)	Default choice: no extra dependencies, fast, scales with data via minibatching. Approximate posterior.
sampler="nuts" (pyro)	Higher-quality posterior samples at (much) higher cost per round. Needs banditry[nuts].

feel_good=True enables Feel-Good Thompson sampling (Zhang, 2021), which reweights the posterior toward optimistic value functions — helpful in contextual settings where plain TS can be insufficiently exploratory.

from banditry import TSConfig, build_agent
agent = build_agent(TSConfig(sampler="langevin", feel_good=True), space)

Rules of thumb

• Start with OFU-GP (surrogate="gp"): strongest baseline at small budgets, no MCMC tuning surface.
• Switch to svgp when rounds get slow from data volume.
• Reach for TS when you want posterior-sampling behaviour (e.g. batch diversity, richer exploration in contextual problems) or a non-GP value function; prefer langevin, and treat nuts as the high-fidelity reference.
• All agents handle mixed spaces and contexts; see Design spaces and Contextual bandits.