You can deploy this model and start earning money today!

Clem2323/
morris-checkpoint

Name: morris-checkpoint
Availability: InStock

Views

No views yet

Champion v2 — Phase 3 v5, step 12000

Source : outputs/2026-05-25/21-02-47/checkpoints/checkpoint_00012000.pt

Architecture : GraphNet 4×128 (215,466 params), aux heads enabled.

Lineage :

Init = champions/v1_phase3_v2_step10000/best.pt (previous champion)
Phase 3 v5 self-play : random_opening K=8 + curriculum (50% random 4v4 starts)
~12k gradient steps before the peak — kept as the strongest MCTS-driven agent

Phase 3 v5 config (key knobs)

network.type            : graphnet 4×128
training.learning_rate  : 3e-5  (warmup 2000 steps, cosine over 60000)
training.warmup_mix_fraction : 0.1  (light minimax-d=5 anchoring)
self_play.asymmetric.prob_random_opening : 1.0
self_play.asymmetric.random_opening_k    : 8
self_play.curriculum.enabled             : true
self_play.curriculum.random_start_fraction : 0.5
self_play.curriculum.pieces_per_player   : 4
self_play.playout_cap.full_sim_fraction  : 0.40
mcts.num_simulations_train               : 800

In-training eval @ step 12000 (n=30 per matchup, opening_random_k=4)

opponent	agent	W	D	L	score
MinimaxAgent(d=3)	net + MCTS200	0.50	0.30	0.20	0.65
MinimaxAgent(d=3)	bare argmax	0.20	0.13	0.67	0.27
MinimaxAgent(d=5)	net + MCTS200	0.15	0.45	0.40	0.38

Compared with the previous champion (v1):

opponent	v1 (step 10k)	v2 (step 12k)
d=3, net+MCTS200	0.63	0.65
d=3, bare argmax	0.72	0.27 ⚠️
d=5, net+MCTS200	0.20	0.38

Trade-off vs v1

v2 is a stronger MCTS-driven agent but a weaker bare prior. It plays better when given thinking time (200+ simulations) but worse if you ever turn the search off. For the web demo (which runs MCTS at every move), that's the right trade-off.

If you ever need a pure policy network with strong argmax behavior (e.g. fast play with zero search), prefer v1.

Notes

30-game in-training eval has Wilson 95% CI ≈ ±0.18, so the +0.02 over v1 is within noise; the meaningful improvement is on d=5 (+0.18) and the fact that the head-to-head vs v1 + 100-game eval should confirm.
Run a proper 100-game eval before treating any margin as decisive.

Clem2323/morris-checkpoint