FRAC-MARL is a decentralised actor-critic multi-agent RL method that tolerates Byzantine edge attacks, where the adversary corrupts messages on communication links. Each agent uses redundancy in two-hop messages to pick out the reliable ones. Under linear value and reward parameterisations, the authors prove the parameters converge almost surely to the attack-free limit points over time-varying graphs, where earlier methods only reached a neighbourhood of them; they also give a topological condition that can be checked in polynomial time, and test on multi-robot formation control.
Relevance
Resilient learning results usually stop at “close to the attack-free answer”. This one claims the exact answer under a specific threat model (the adversary sits on the communication links and the agents themselves are honest), which makes it a good test of which Byzantine-robust learning guarantees survive moving robots and changing graphs in robust robotics.
Links
- arXiv
- Subfield: Multi-robot systems, Robot learning