This directory contains exploratory work and experiments for Infra-Bayesian Reinforcement Learning Agents Outperform Classical RL For Worst-Case Robustness. For review or reproduction, start with the entries below rather than the individual notebooks, which may contain exploratory work.
Run commands from the repository root after installing the project dependencies.
| item in Infra-Bayesian Reinforcement Learning Agents Outperform Classical RL For Worst-Case Robustness | experiment | command | outputs |
|---|---|---|---|
| Classical-equivalence validation figure | experiments/fllor2 |
uv run python -m experiments.fllor2.validate_classical |
experiments/fllor2/validate_classical.png |
| Knightian-uncertainty worst-case robustness figure | experiments/fllor2 |
uv run python -m experiments.fllor2.ku_demo |
experiments/fllor2/ku_demo_outputs/ |
| Newcomb predictor-accuracy figure | experiments/fllor2 |
uv run python -m experiments.fllor2.newcomb_accuracy |
experiments/fllor2/newcomb_accuracy_outputs/ |
| Trap-bandit robustness figures and summary table | experiments/alaro/trap_bandit |
uv run python -m experiments.alaro.trap_bandit.run |
experiments/alaro/trap_bandit/results_report_200_pcat001/ |
The output files and directories listed above are checked in with the figures and summaries used by Infra-Bayesian Reinforcement Learning Agents Outperform Classical RL For Worst-Case Robustness. Re-running the commands regenerates those artifacts with the settings used for that work.