gittins

An opinionated, highly optimised contextual bandit engine.

• Online by nature. Learns one observation at a time, in O(1) work and fixed memory, as long as open decisions are regularly resolved.
• Non-stationarity is expected. The engine adapts as the relationship between context and feedback drifts, and never learns anything it cannot eventually unlearn.
• Dynamic actions and context. The candidate set can change on every call; you never declare the number of actions up front.
• Safe reward handling. Rewards may arrive late or never. Constructing invalid training data is made hard by design.
• Speed and determinism. Bit-identical results across platforms and language bindings, enforced by a golden test corpus.
• Bring your own model. Swap in your own scoring or exploration and inherit everything else.

Full documentation and user guide: docs.getgittins.dev

Install

pip install gittins

Usage

The state is an opaque handle updated in place — calls return only their result, and every alias of the handle observes the current state. Persist it with serialize, which returns one plain string.

import time
import gittins

state = gittins.create(bits=8, horizon=3600.0)  # model size, seconds to resolve

candidates = [
    ("banner-sale", {"discount": 0.2}),
    ("banner-new", {"discount": 0.0}),
    ("banner-plain", {}),
]
context = {"device": "mobile", "hour": 14}

record = gittins.decide(state, context, candidates, time.time(), "web-1")
arm_id = candidates[record.chosen][0]
# ... your code: act on the choice, then report the outcome

gittins.learn(state, record.decision_id, 1.0, time.time())
gittins.expire(state, time.time())  # resolve anything past its horizon

open("bandit.txt", "w").write(gittins.serialize(state))

Feature values are typed by what you pass: strings are categorical, ints, floats and bools numeric, None absent. Anything else raises ValueError naming the feature.

Bringing your own model

decide takes optional score and explore callbacks, and learn/expire take train. Each crosses the Python/Rust boundary once per call — with all candidates, all estimates, or the one resolved record — never once per candidate, so the one boundary crossing per decision is preserved.

Offline policy evaluation

log_line renders a decision record or resolution as one canonical experience-log line. Append it verbatim; it is exactly what the gittins CLI's verify / eval / sweep / replay consume.

with open("decisions.jsonl", "a") as log:
    log.write(gittins.log_line(record) + "\n")

API

License

MIT. Named after John Gittins, whose index (1974) established that exploration has a precise, computable value.