v1.0.0

LightRL 1.0.0

Lightweight multi-armed bandit library for Python. Zero heavy dependencies. Built for agents.

Highlights

• 7 bandit strategies: Epsilon-Greedy, Epsilon-First, Epsilon-Decreasing, UCB1, Thompson Sampling, GreedyBanditWithHistory, LinUCB
• BanditRouter: manage multiple named decision points with one object
• LinUCB contextual bandit: context-dependent decisions with pure-Python linear algebra
• JSON persistence: save/load on all bandits and the router
• Warm start: initialize with prior beliefs
• EMA decay: make any bandit adaptive to non-stationary environments
• Python 3.8 - 3.14 support

Examples

17 runnable examples covering A/B testing, ad serving, LLM cost optimization, agent fleet ROI, dynamic pricing, database query routing, and more.

Install

pip install lightrl

v0.1.1

Publishing documentation.

v0.1.0

Minimal working version.

What is new:

• Bandit
• EpsilonGreedyBandit
• EpsilonFirstBandit
• EpsilonDecreasingBandit
• UCB1Bandit
• GreedyBanditWithHistory
• two_state_time_dependent_process

v0.0.1

This is the first release.