| name | ix-bandit |
| description | Multi-armed bandit simulation — epsilon-greedy, UCB1, Thompson sampling |
| disable-model-invocation | true |
Multi-Armed Bandits
Simulate bandit algorithms to find the best arm given uncertain rewards.
When to Use
When the user needs to compare exploration-exploitation strategies, simulate A/B testing, or understand bandit algorithms.
Capabilities
- Epsilon-greedy — Explore with probability ε, exploit otherwise
- UCB1 — Upper Confidence Bound (optimism in the face of uncertainty)
- Thompson sampling — Bayesian approach with posterior sampling
Key Concepts
- Higher true mean → algorithm should pull that arm more
- UCB1 gives logarithmic regret bounds
- Thompson sampling is often best in practice
Programmatic Usage
use ix_rl::bandit::{EpsilonGreedy, UCB1, ThompsonSampling};
MCP Tool
Tool name: ix_bandit
Parameters: algorithm, true_means, rounds, epsilon