Authors:
Mathis Delsart, Achille Morenville, Eric Piette

Venue:
IEEE Conference on Games (CoG), 2026 (accepted)

Topics:
deep reinforcement learning, real-time strategy games, MicroRTS, entity-based reasoning, spatial reasoning, U-Net, Transformer

Links: PDF

In MicroRTS, a real-time strategy benchmark, winning requires long-range coordination between many units. Current deep reinforcement learning agents commonly reason over spatial feature maps and treat units implicitly through stacks of channels.

This paper introduces UECD, a hybrid architecture that explicitly couples spatial and per-unit reasoning through a multi-scale convolutional backbone and a Transformer over unit entities. It also presents a PPO-based training recipe derived from systematic ablation.

On the basesWorkers16x16A map, UECD outranks prior competition winners in an open-source and reproducible tournament, reaching a 96.67% win rate.

MicroRTS is a compact but challenging benchmark for real-time strategy AI, combining large action spaces, sparse long-horizon rewards, multi-unit coordination, and strict real-time constraints.

The proposed architecture decomposes RTS reasoning into complementary components: spatial perception through a CBAM-gated U-Net, relational reasoning through an entity Transformer, and global spatial reasoning through self-attention over the bottleneck grid.

The results show that combining spatial and entity-based representations can improve competitive MicroRTS play at modest compute cost, while also highlighting the limits of single-map specialization under layout and scale shifts.

Delsart, M., Morenville, A., Piette, E. (2026). Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers. IEEE Conference on Games (CoG). Accepted.

@inproceedings{delsart2026combining,
  author    = {Delsart, Mathis and Morenville, Achille and Piette, Eric},
  title     = {Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers},
  booktitle = {IEEE Conference on Games (CoG)},
  year      = {2026},
  note      = {Accepted}
}