Multi-unmanned aerial vehicle (UAV) platforms integrate radio-frequency (RF) sensing, datalinks, and onboard embedded compute; adversarial electronic warfare (EW) degrades these subsystems through jamming and forces decentralized control policies to act on fragmented observations—a setting aligned with intelligent electronic systems and autonomous robotics in contested spectrum. Cooperative swarms then face two compounding failure modes: loss of coherent situational awareness, and reward-driven passive survival that suppresses mission completion. Memory-based multi-agent reinforcement learning (MARL) partially addresses the first but tends to reinforce the second; dense intent shaping addresses the second but becomes unreliable when observations are incomplete. We propose CIDA (Cognitive–Intent Dual-Stream Architecture), a reinforcement learning framework that decouples belief reconstruction from tactical intent at the representation level while coupling them through a unified actor–critic update. The cognitive stream encodes a 64-step observation history with a pre-normalized Transformer to reconstruct threat belief; the intent stream supplies a hierarchical potential field (reconnaissance, threat-weighted engagement, and approach incentives). A steady-state training mechanism (dynamic reward scaling and adaptive gradient clipping) stabilizes Transformer-based on-policy learning under non-stationary multi-agent dynamics. In a complex terrain scenario with SAM, AAA, and jammer assets, CIDA reaches 96.15% task success versus 12.21% (memoryless PPO) and 25.28% (MAPPO+RNN), with ablations showing nonlinear coupling and emergent tactics such as jammer bypass and weak-sector traversal. Results are robust to a four-fold sweep of the intent-shaping weight (above 90% success).
Chen et al. (Sat,) studied this question.