Methodology

Base system

The benchmark uses an industrial-flavored multi-agent process-control environment: the MARL-SortingEnv recycling line, which combines SortingEnv and ContainerGym (see Base environment).

Fault injection

Faults are introduced through a wrapper around the environment's step() process. The injection layer is configuration-driven and reproducible from a seed.

Recovery layer

Four recovery mechanisms are implemented as swappable modules above the same base MARL system. This preserves the apples-to-apples comparison.

Evaluation

Each mechanism × fault-type combination is evaluated across 8 random seeds (and 3 training seeds for the two trained mechanisms), with 95% confidence intervals.