New MARL framework slashes energy waste in relay networks by 95%
Multi-agent deep Q network hits near-optimal power control with only outdated channel data...
In this paper, the authors tackle a cooperative game in two-hop relay networks where each relay autonomously maximizes energy efficiency. The key challenge is that instantaneous channel state information (CSI) is hard to obtain in practice—only outdated, partial CSI is available. To address this, they define a delayed reward-based state-action value function and build a multi-agent deep Q-network (MADQN) learning framework. This allows relays to coordinate without perfect real-time channel knowledge, using historical data to guide decisions.
Simulation results demonstrate that their method considerably outperforms potential alternatives (e.g., Q-learning with full CSI, distributed heuristics) and is merely 5.2% away from the theoretical optimal solution computed with perfect CSI. The authors analytically prove that game-theoretic approaches with instantaneous CSI provide upper bounds for their proposed method, confirming the framework's near-optimality. This work is significant for real-world deployments where perfect channel knowledge is unavailable, enabling substantial energy savings in wireless relaying systems with minimal performance loss.
- Uses a cooperative multi-agent deep Q-network with delayed rewards to handle outdated CSI
- Achieves energy efficiency within 5.2% of the theoretical optimal solution
- Outperforms all alternative methods in simulations with partial and imperfect channel knowledge
Why It Matters
Enables near-optimal power control in relay networks without perfect channel data, cutting energy costs in real-world deployments.