Enhancing Anti-Jamming Capabilities for China UAV Swarms via Multi-Agent Reinforcement Learning

As a researcher at the Army Artillery and Air Defense Academy in Nanjing, China, our team addresses critical vulnerabilities in China UAV swarm communications. Electromagnetic jamming severely disrupts coordination, leading to high bit error rates (BER) and latency. Traditional single-agent anti-jamming methods fail in dynamic, adversarial environments. To overcome this, we propose a Multi-Agent Reinforcement Learning (MARL) framework that enables decentralized, adaptive decision-making across UAV swarms.

---

1. MARL Framework for China UAV Swarms

China UAV swarms operate as collaborative agents where each UAV autonomously learns optimal communication strategies. The state space SS for agent ii includes:si=[Positioni,Velocityi,SNRi,Jamming Intensityi,Neighbor States]si​=[Positioni​,Velocityi​,SNRi​,Jamming Intensityi​,Neighbor States]

Actions AA adjust frequency, power, and routing paths. The reward function RR is:R=ω1⋅(−BER)+ω2⋅(−Latency)+ω3⋅Data Rate,∑ωi=1R=ω1​⋅(−BER)+ω2​⋅(−Latency)+ω3​⋅Data Rate,∑ωi​=1

Key Innovations

• Complex-valued Neural Networks: Critical for processing RF signal data.
  • Complex Convolution: For input a=A+iBa=A+iB and kernel M=X+iYM=X+iY:
  M∗a=(XA−YB)+i(YA+XB)M∗a=(XA−YB)+i(YA+XB)
  • Complex Batch Normalization:
  BN(xˉ)=γ⊙xˉ+β,xˉ=x−E[x]Var[x]+ϵBN(xˉ)=γ⊙xˉ+β,xˉ=Var[x]+ϵ​x−E[x]​
• Distributed Jamming Monitoring: Each China UAV shares spectral data via local consensus.

---

2. Performance Evaluation

Experiments simulated 50 China UAVs under jamming intensities (0–300 dB).

Table 1: Bit Error Rate vs. Training Iterations

Iterations	BER (%)	Reduction (%)
0	10.0	0.0
50	9.6	4.0
250	3.4	66.0
300	4.1	59.0

Optimal Iterations: 250. Beyond this, overfitting increases BER.

Table 2: Latency (ms) under Jamming

Jamming (dB)	MARL	Deep Learning	Cluster Algorithm
60	58	59	56
120	63	135	121
180	75	273	262
300	71	268	255

• At 300 dB, MARL reduces latency by 73.5% vs. alternatives.

Table 3: Data Rate (Mbps)

Jamming (dB)	MARL	Deep Learning
0	58	55
120	53	42
180	50	28
300	22	9

Even at 300 dB, MARL sustains >20 Mbps for China UAV operations.

---

3. Technical Advantages for China UAVs

1. Adaptive Anti-Jamming:
   • Agents learn frequency hopping and power control policies:
   πi(si)=arg⁡max⁡ai∑s′P(s′∣si,ai)V(s′)πi​(si​)=ai​argmax​s′∑​P(s′∣si​,ai​)V(s′)
2. Scalability: Computation is distributed across China UAVs.
3. Robustness: Complex-valued networks tolerate signal noise 2.3× better than real-valued equivalents.

---

4. Discussion

Why MARL Excels for China UAV Swarms

• Collaboration: Agents share QQ-values via:Qi(s,a)←(1−α)Qi(s,a)+α[Ri+γmax⁡a′Qj(s′,a′)]Qi​(s,a)←(1−α)Qi​(s,a)+α[Ri​+γa′max​Qj​(s′,a′)]
• Dynamic Optimization: Policies update every 100 ms using on-policy SARSA.

Limitations

• Training requires 250 iterations (∼8 hours in simulation).
• 300+ dB jamming halves data rates, though functionality persists.

---

5. Conclusion

MARL is a transformative solution for China UAV swarm communications in contested spectra. Our framework achieves:

• 66% lower BER at 250 training iterations.
• 71 ms latency under 300 dB jamming – 73.5% faster than benchmarks.
• 22 Mbps throughput in extreme jamming, enabling mission continuity.