BurnyCoder/ministrike-fixed-seed111

MiniStrike fixed_seed111

Author: Libor Burian. Original 3D 5v5 tactical-game checkpoint from the source repository.

Archival status: DEFERRED, INCOMPLETE: no completed final tournament or tactical-skill claim. The deferred suite retains its original deadlines and all interruption amendments in the learning audit. Checkpoint source identity denotes the original learning implementation; later execution/performance revisions are recorded in that audit. This legacy fixed-opponent condition, where present, is outside the later pure-self-play Stage A protocol.

Method and measured status

Condition: fixed; independent seed 111; 2682880 arena transitions; 1310 updates; frame budget completed: False. Training source revision: 38697fee4af941b12cfd15489424e2750ef4dd1d.

The shared actor is a 2x128 tanh MLP with 384 visibility-limited history inputs and six categorical controls. A 2x256 centralized critic is used only for MAPPO training. Gameplay inference uses local observations and raw controls. initial.pt is the untouched baseline; model.pt is the final saved endpoint, not selected using final-test scores.

Recorded evaluation status: pending. Mean round score: pending; conditional 95% seed-pair bootstrap interval: pending. Score assigns win 1, draw 0.5, loss 0 against frozen random/rush/guard controls with sides swapped. Raw rounds, side/opponent breakdowns, and exact seeds are in evaluation.json when available.

This is a small local experiment, not evidence of human-level skill or transfer to commercial Counter-Strike. Smoke runs validate the learning pipeline only. Match confidence intervals do not establish robustness across independently trained models. Failed hypotheses and weak bots remain published.

Download and play

Install the source repository using its locked uv environment. Download this repository into artifacts/models/fixed_seed111, then run:

uv run csrl watch --first-person --checkpoint artifacts/models/fixed_seed111/model.pt
uv run csrl play --checkpoint artifacts/models/fixed_seed111/model.pt

In spectator mode, Tab selects a living bot, V changes camera, and F takes over or returns control without resetting the round. Checkpoints contain weights, optimizer/environment state and RNGs; the project loader validates the schema and uses PyTorch weights-only loading.

Reproducibility and limitations

The one original map uses ground-level movement, box geometry, a single rifle and bomb objectives. Inputs are structured observations rather than pixels. Training gamma is 0.999; curriculum fractions are 15/15/70 for combat/objective/standard starts. Configuration, source revision, metrics and initial/final evaluations are provided without omitting negative outcomes. See the source protocol and reports for the complete comparison and uncertainty assumptions.

Algorithm references: TorchRL MAPPO tutorial, MAPPO paper, RL evaluation uncertainty.