Resource-Aware Strategic Planning for Partially Observable Star-Sector Tactical Games

Autonomous agents in adversarial environments must allocate sensing resources, deploy spatial assets, and control combat actions under partial observability. We formulate this coupling as a three-phase partially observable decision process with heterogeneous action spaces. Strategic Planner provides a structured, non-recurrent policy that combines a shared phase-conditioned representation, masked entity attention, a six-channel tactical-map encoder, static plan tokens, competitive entity–map weighting, and phase-specific heads. Configuration uses a probability-consistent tanh-squashed Gaussian for 11 bounded controls and three categorical choices; deployment and battle use legality-masked categorical distributions. We evaluated the architecture against sequence-aware PPO-GRU and non-recurrent baselines under matched environment-transition budgets. Across eight training seeds, the deterministic win rate was 21.8 ± 28.0% for Strategic Planner and 25.8 ± 22.7% for PPO-GRU; their paired difference was −4.0 percentage points (95% bootstrap interval [−21.8,18.1]). Strategic Planner had numerically close mean episode reward (78.0 versus 77.5) and lower mean Command Point cost (753.5 versus 775.9), but paired intervals do not establish an advantage on these endpoints. Peak batch-one GPU allocation was 49.1 versus 54.6 MiB. Three-seed component ablations were inconclusive, and shifted-map outcomes varied. The study delivers an auditable, probability-consistent phase-coupled architecture and a quantified account of its trade-offs with recurrent memory.