Anti-Jamming Algorithm Based on Stackelberg Game and Soft Actor-Critic

This paper studies a hybrid anti-jamming architecture using a Stackelberg game as its theoretical formulation in which a Soft Actor-Critic (SAC) agent selects a robustness radius and a power-control coefficient, while model-based power allocation and second-order cone programming determine transmit powers and receive-combining weights. The experimental evaluation uses a frequency-selective multipath Rician channel with diffuse scattering, path delays, and Doppler evolution. Three training runs are assessed using deterministic policies on fixed control channels and a separate test set. On the separate test, training increases mean reward from 4.133 to 14.483 and reduces average-SINR outage at 12 dB from 0.935 to 0.403, accompanied by an increase in normalized transmit power from 0.690 to 0.979. An independently tuned fixed-parameter baseline achieves reward 14.923 and outage 0.374. The results demonstrate improvement over the initial policy, but do not establish an advantage over tuned fixed parameters. Robust feasibility remains conditional on the channel-uncertainty set and numerical feasibility; the simulated jammer uses an exogenous allocation rule, so the experiments do not verify a Stackelberg equilibrium.