v0.1.0 - Deep Reinforcement Learning for Competitive Agents in MicroRTS: Architecture, Training, and Tournament Evaluation

Master's thesis release, UCLouvain, 2026.

This is the v0.1.0 archival release of the thesis pipeline:

• UECD-Best tops a 19-agent IEEE-CoG-style round-robin at 96.67% pool win rate.
• Beats RAISocketAI (IEEE-CoG competition winner) in 65.7% of head-to-head games.
• Trained on 9.47 GPU-days / 350M steps (vs RAISocketAI's 23.6 GPU-days / 500M).
• UECD-MultiMap generalises across 5 layouts of 3 different sizes with no per-map collapse.

The full pipeline (training, evaluation, tournament, behaviour cloning, benchmarks, analysis) is reproducible from one of two automated setup scripts under setup/ (laptop or CECI HPC).

See dissertation/dissertation.pdf for the full thesis and CITATION.cff for citation metadata.

A Zenodo DOI will be assigned automatically once the Zenodo integration archives this tag (10-30 minutes after release).

TensorBoard event archives

Per-agent TensorBoard event files for every agent shipped under
data/agents/.

These were kept out of the repo (~4.4 GB total) because git is the wrong
vehicle for big append-only binary streams. They live here as release
assets instead: free, no LFS quota, downloadable individually.

What's in the archive

Each file is the events.out.tfevents.* log of one agent, renamed for
clarity. Naming convention: <agent-dir-name>_<step-range>.tfevents.

Asset	Size	Agent dir under data/agents/
UECD-SingleMap-Best_0-360M.tfevents	1.2 GB	UECD-SingleMap-Best/ (three-phase fine-tune log: Phase 0 + 1 + 2 cat-merged into one continuous TFRecord stream, readable by TensorBoard with two normal phase-boundary monotonicity warnings)
UECD-SingleMap-Rushed_0-300M.tfevents	704 MB	UECD-SingleMap-Rushed/ (= PhasedRL-300M, the rush-collapse demo run)
UECD-MultiMap-Best_0-330M.tfevents	711 MB	UECD-MultiMap-Best/
GridNet-SingleMap_0-300M.tfevents	588 MB	GridNet-SingleMap/ (the published baseline)
UECD-MultiMap_0-200M.tfevents	580 MB	UECD-MultiMap/
UECD-BC-PPO_0-100M.tfevents	229 MB	UECD-BC-PPO/
UECD-SingleMap-TopFeats_0-100M.tfevents	207 MB	UECD-SingleMap-TopFeats/
UECD-SingleMap-AllFeats_0-100M.tfevents	206 MB	UECD-SingleMap-AllFeats/

UECD-BC is the only agent without an entry: its training script logs
to stdout only, no TensorBoard events are written.

How to use

Download one file:

gh release download tfevents-agent-archive -p UECD-SingleMap-Best_0-360M.tfevents

Or grab the lot:

gh release download tfevents-agent-archive

Then point TensorBoard at the directory you downloaded them to:

tensorboard --logdir .

Growing the archive

The archive is frozen at the 8 agent-level files above. Per the user
decision documented in PR #56, the arch/feat ablation tfevents
(~3.9 GB of additional uploads) were deliberately not added: the
eval/results.csv published under data/ablation/{arch,feat}/eval/
covers the publishable comparison, and intermediate training curves at
the 100 M ablation budget are not worth the release-storage footprint.

RAISocketAI wheel (rl_algo_impls v0.2.1)

Bundled redistribution of the RAISocketAI competition bot wheel
(rl_algo_impls v0.2.1, ~225 MB) attached as a release asset so a fresh
clone of this repo can fetch it deterministically. Original source:
https://github.com/sgoodfriend/rl-algo-impls (MIT licence).

Used internally by setup/local.sh and setup/cluster.sh to install the
RAISocketAI tournament bot (pip install --no-deps).

SHA-256: 1e0a60133f4b96fa95f4331e258fd20495d2209d88c319116ac1bd19431e71d1