Hannah B. Pasandi

Cascade Safe Learning

Learning systems with decision-time safety: constraint masks, phase-transition theory, and zero-violation guarantees for shared infrastructure.

Cascade Safe Learning

NodeZero

A per step safety layer for learned Wi Fi control.
zero propagating cascades10.4 million decisions0.3 ms per decision

NodeZero masks cascade inducing actions before a learned controller acts. It treats interference as an epidemic and uses a checked oracle so dense Wi Fi can keep learning gains without unsafe decisions.

mask the cascade-inducing action before it actscontrollerlearned policymaskoracleWi-Fi actionsafecascade action blocked
Cascade Safe Learning

BlackWidow

Safe reinforcement learning for zero violation Wi Fi.
47.2 Mbps throughputzero propagating violations47 node testbed

BlackWidow removes unsafe wireless actions before policy execution. It uses a feasibility oracle so high throughput learning never assigns probability to cascade inducing configurations.

never assign probability to an unsafe actionpolicy distribution✗→ 0feasibilityoraclezero-violation execution
Cascade Safe Learning

INTACT

Phase transition theory for interference cascades.
R0 17.4 at peak94% outbreak probabilityzero violation mask

INTACT formalizes dense Wi Fi interference as an epidemic process with a threshold at R0 equals one. It explains why expected cost safe RL fails when one violation can spread network wide.

interference as an epidemic with a thresholdR0R0 = 1containedoutbreakexpected-cost safe RL fails when one violation spreads
Cascade Safe Learning

VETO

Verified safe reinforcement learning with constraint filtering.
zero violation objectiveconstraint predictoradaptive threshold

VETO filters unsafe actions before environment execution using calibrated uncertainty and threshold optimization. It targets zero violations while preserving reward in safety critical domains.

filter unsafe actions before the environmentpolicyconstraintcalibratedexecute ✓veto ✗adaptive threshold targets zero violations