Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers

Delsart, Mathis;Morenville, Achille;Piette, Eric
(2026) IEEE Conference on Games 2026 — Location: Madrid, Spain (1.September.2026)

Files

cog26.pdf
  • Open Access
  • Adobe PDF
  • 1.05 MB
Download

Details

Authors
Abstract
In MicroRTS, a real-time strategy benchmark, winning requires long-range coordination between many units. Yet, current deep reinforcement learning agents reason only over spatial feature maps and treat units implicitly through stacks of channels. The contributions of this paper are twofold: (i) UECD, a hybrid architecture coupling spatial and per-unit reasoning explicitly via a multi-scale convolutional backbone and a Transformer over unit entities; and (ii) a PPO-based training recipe derived from systematic ablation. On the basesWorkers16x16A map, UECD outranks prior competition winners, topping an open-source and reproducible tournament with a 96.67% win rate.
Affiliations

Citations

  • APA
  • Chicago
  • FWB

Delsart, M., Morenville, A., & Piette, E. (2026). Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers. In Rui Prada, Joana Brito (ed.), IEEE Conference on Games 2026. https://doi.org/10.1007/3-540-56393-8_38