Rank-Based Softmax Ensemble of DQN and PPO for Adaptive Anti-Jamming Wireless Transmission

Abstract

Jamming attacks pose a critical threat to wireless communication systems, necessitating intelligent, adaptive countermeasures. This paper proposes a training-free Rank-Based Softmax Ensemble that fuses pre-trained Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) agents at inference time for adaptive anti-jamming transmission and provides a rigorous comparative study of all three methods. The problem is formulated as a Markov Decision Process (MDP) with a 1,344-state space that captures jamming power, data queue occupancy, and energy buffer status. A normalized multi-objective reward incorporating a structural penalty Ωt eliminates deadlocks and neutral-action inflation. DQN and PPO are trained under identical configurations; the ensemble requires no training, dynamically adjusting fusion weights via credit-tracked softmax over a sliding lookback window. PPO achieves the highest overall return (≈585.1 balanced and ≈697.4 high throughput) and the fastest convergence. The proposed ensemble demonstrates competitive performance in high throughput (≈693.2) and outperforms both standalone models in the default and energy efficiency configurations (557.6 vs. PPO 551.7, DQN 520.7, and 530.2 vs. PPO 527.3, DQN 489.2, respectively) and a corresponding energy efficiency (η=56.69 vs. PPO 16.27, DQN 4.50), with the highest update rate (520 updates/sec), lowest inference latency (0.53 ms), and strongest jamming resilience, all at zero training cost. Practical algorithm selection guidelines are provided.