RL-Sim — Reinforcement Learning Simulation

Home for experiments that simulate intervention policies before we run real-world pilots. The repo will host:

• Environment definitions (gym-like) for different study designs.
• Policy training scripts and evaluation notebooks.
• Reporting utilities that summarize safety signals for the ethics board.

TL;DR

• python scripts/demo.py — synthetic rewards, CSV summary, HTML report (outputs/demo_*)
• python scripts/plot_severity.py --output docs/severity_mean_rewards.png
• python scripts/doctor.py — env check (Python, deps, writable outputs)
• Install: python3 -m venv .venv && source .venv/bin/activate && pip install matplotlib

See ROADMAP.md for the current build order and meta/launchpad/project-priorities.md for near-term deliverables.

Current Prototype

The src/rl_sim/bandit.py module now exposes stable and volatile reward schedules plus a minimal epsilon-greedy simulator. Start with build_default_bandit() in a notebook to compare conditions before wiring up heavier experiments.

Comparing Severity Groups

Use simulate_severity_groups() with the built-in LOW_SEVERITY and HIGH_SEVERITY profiles to generate quick contrasts before building plots. Example:

from rl_sim import simulate_severity_groups, LOW_SEVERITY, HIGH_SEVERITY
results = simulate_severity_groups(trials=200, profiles=[LOW_SEVERITY, HIGH_SEVERITY], rng_seed=7)
print(results['low']['mean_reward'], results['high']['mean_reward'])

Swap in your own SeverityProfile instances if you need different schedules or exploration rates.

Figure Workflow

Install matplotlib (pip install matplotlib) and run python scripts/plot_severity.py to produce docs/severity_mean_rewards.png, our first sketch for the simulation paper. Details live in docs/figures.md.

Documentation

• docs/quickstart.md — one-page demo/plot/doctor steps.