This algorithm reaches Byzantine-resilient vector consensus without being told how many neighbours might be faulty. Each agent scores its neighbours from observed state discrepancies, accumulates that evidence into reputation weights, and can push a neighbour’s weight to exactly zero. The weights go straight into the consensus update, per-agent cost stays low for high-dimensional states, and experiments compare it with classical resilient consensus across attack types and state dimensions.
Relevance
W-MSR style methods need a bound F on faulty neighbours and a graph robust enough for that F, and in the field F is rarely known in advance. Dropping the bound is attractive; an open question is what a patient attacker, one that behaves well before it lies, can do to the reputation scores. See resilient consensus and robust robotics.
Links
- arXiv
- Subfield: Multi-robot systems, Security