BE-CTDE: Blockchain-Enabled Multi-Agent Reinforcement Learning with Anomaly Detection and Token Economics
- Published
- Source
- arXiv
- Paper number
- 1373
- Field
- AI / Agents
- arXiv ID
- 2610.34121
Key points
- The authors proposed BE-CTDE, which turns blockchain from a passive record into an active training-management layer where behavioral evidence changes learning influence.
- A behavioral verifier compares executed trajectories with shared-policy recommendations, and its flags drive a token mechanism combining slashing, burning, rewards, and compensation.
- They showed perfect detection is unnecessary: with conditional detection probability above 1/3, an adversary is excluded from training in finite expected time.
- In multi-agent path finding experiments, success rates under attack were substantially higher than the state-of-the-art SCRIMP planner, even with an adversarial majority.
- Collision rates were up to 35% lower than detection without economic consequences and up to 43% lower than conventional stake slashing, and the 8-agent setting reached 100% success rate at every obstacle density.
Paper links
External research summaries. These are not HDATF publications or measured product results.