Surriworks / ModelsDownload evidence

Space Mining evaluations

Loading the recorded evaluations…

Against Hard AI

Hard = 1,000

Ratings compare placement against the shipped Hard bot. Small samples are provisional; missing evaluations stay unrated.

Frozen model results for the selected game format
ModelRelative Elo*Wins / gamesMean rankMapsEvidence

Next experiment

Same teacher, replay and held-out maps. Change one architecture decision at a time.

  1. CNN–TransformerBaseline · one encoding layer, three reasoning layers
    Not rated
  2. Gated HopfieldSame backbone · 256 learned memory patterns
    Not rated
  3. Wider Hopfield192-wide backbone · test whether more capacity helps
    Not rated

New architectures are experimental. No cloud training result is implied by these entries.

The decision path

Public board + ships + action history
CNN + Transformer encoder
Optional Hopfield + residual gate
3 Transformer blocks → policy & value
Candidate scoring → action

Planned extension: simulate candidate sequences against history-conditioned opponent policies in sampled hidden worlds. Full game search is not implemented yet.

How to read the ratings

*A regularized pairwise estimate on the Elo scale, not yet a multi-opponent league rating. Each independent map contributes one mean score; repeated seat rotations do not become independent maps. A small smoothing prior keeps all-loss samples finite. Ratings from different formats should not be compared directly.

First place receives 4 tournament points and second receives 2. Lower-place points remain unconfirmed, so tournament-point totals are not displayed. Evaluation runs accelerate wall time while retaining the live action-count budget; they do not certify network latency.