Shielding¶
MASA provides tutorials for both risk-budget-based probabilistic shielding and winning-region safety-game shielding.
Winning-region Mini PacMan example¶
The winning-region safety-game shielding guide explains support-based synthesis and the preemptive and postposed interfaces.
The companion Mini PacMan notebook uses the safety property
and compares:
- mask-aware preemptive action selection;
- postposed uniform-random replacement of unsafe proposals.
The notebook is a small shielding-interface demonstration rather than a training benchmark or a guarantee of task completion.
Open tutorial/05_minipacman_deterministic_shielding.ipynb
It uses replacement_strategies.random_safe. The replacement random generator is
independent of the environment seed.
Probabilistic shielding tutorials¶
These tutorials show how MASA's probabilistic shielding wrappers turn risk-bound analysis into environment interfaces that reinforcement-learning algorithms can consume.
- Probabilistic Shielding MiniPacman
builds a PCTL-constrained MiniPacman environment, wraps it with
ProbShieldWrapperDisc, and inspects safety bounds, successor dynamics, and projected safe actions. - Safety Abstractions Pacman Coins shows why structured Pacman coin observations need a discrete safety abstraction before probabilistic shielding can compute safety bounds.
- FrozenLake Shielding shows how the same discrete shield-synthesis path applies to Gymnasium-style FrozenLake dynamics.
Runnable notebooks: