UC-PSRO: Utility-Conditioned Policy-Space Response Oracles with a Communication-Dropout Curriculum for Game-Theoretic Course-of-Action Generation in Adversarial Swarms
Phillip Jiang
We present UC-PSRO, a utility-conditioned policy-space response oracles framework for generating game-theoretic courses of action in adversarial drone swarm operations under communication-limited conditions. The framework combines self-play policy optimization, weighted utility conditioning for adaptive policy steering, and a communication-dropout curriculum that trains robustness to progressive comms failure. On a maritime swarm scenario spanning 25–200 agents, the communication-dropout curriculum alone proved the most reliable component — counterintuitively improving success from 35% to 62% as communication disruption intensified — while combining all three components slowed convergence within a fixed training budget without a corresponding gain in exploitability over fixed-opponent baselines.
HDST-GNN: Heterogeneous Dynamic Spatiotemporal Graph Neural Networks for Multi-Object Tracking in UAV Aerial Imagery
Phillip Jiang
We present HDST-GNN, a heterogeneous dynamic spatiotemporal graph neural network for multi-object tracking (MOT) in UAV aerial imagery. The model represents each frame as a heterogeneous graph with three node types — new detections, confirmed tracklets, and lost tracklets — and builds edges adaptively based on estimated camera altitude. An occlusion-gated temporal aggregation mechanism prevents occluded objects from corrupting neighboring trajectory representations. HDST-GNN achieves 94.51% MOTA and 97.24% IDF1 on standard UAV MOT benchmarks, reducing identity switches by 81% over the SORT baseline.
RelGT-AC: A Relational Graph Transformer for Autocomplete Tasks in Relational Databases
Phillip Jiang
We introduce RelGT-AC, a relational graph transformer designed for autocomplete prediction tasks in relational databases. The model constructs a heterogeneous graph over database tables and rows, encodes text-heavy features via a pretrained language model backbone, and applies graph transformer attention across relational edges. RelGT-AC achieves up to +10 AUROC over GraphSAGE baselines on standard relational database ML benchmarks, demonstrating that graph transformer architectures generalize effectively to structured tabular domains.