Contextual bandits

In a contextual bandit, part of the input is observed, not chosen: each round the environment hands you a context, and the agent optimises the remaining parameters conditionally on it.

In banditry, contexts are ordinary parameters of the design space that you pin each round via the second argument of suggest:

import numpy as np
from banditry import DesignSpace, OFUGPConfig, build_agent

space = DesignSpace.parse([
    {"name": "x0", "type": "num", "lb": -1, "ub": 1},   # decision variable
    {"name": "x1", "type": "num", "lb": -1, "ub": 1},   # observed context
])
agent = build_agent(OFUGPConfig(rand_sample=4, surrogate="gp", noise_std_proxy=1.0), space)

for _ in range(30):
    context = {"x1": observe_environment()}   # whatever the world gives you this round
    rec = agent.suggest(1, context)           # x1 is pinned; x0 is optimised given x1
    y = np.asarray(objective(rec), dtype=float).reshape(-1)
    agent.observe(rec, y)

The pinned columns are fixed both during quasi-random warmup and during model-based optimisation (the acquisition is optimised only over the free parameters).

Best observation for a context

get_best_id accepts the same fix_input dict and returns the best matching observation (falling back to the global best if nothing matches):

best_for_context = agent.get_best_id(fix_input={"x1": 0.7})

Notes

• Any subset of parameters can serve as context, including categorical ones.
• The surrogate is trained on the full parameter vector (decisions + contexts), so information is shared across contexts — nearby contexts inform each other.
• Both agent families support contexts (support_contextual = True).