Karma economies get model-free RL: DQN agents learn equilibrium without full knowledge
Deep Q-networks let agents learn fair resource allocation in large populations without knowing the game model.
Deep Dive
Cederle, Bolognani, and Susto study model-free equilibrium learning in Karma economies — a class of fair non-monetary resource allocation mechanisms. First, they show a single agent can
Key Points
- Single agent can join an existing Karma equilibrium and learn via DQN with suboptimality bound O(1/√N_s) + O(1/N)
- Full population can learn Nash equilibrium from scratch using deep RL + fictitious play + smoothed policy iteration
- Karma economies are non-monetary, fair resource allocation mechanisms applicable to autonomous vehicles, IoT, and shared computing
Why It Matters
Enables fair, large-scale resource allocation without central authority or game model – a key step for decentralized AI systems.