V7 — AlphaStar (scaled search)
V7 keeps V6's discounted value target and simply leans into search. AlphaGo Zero's strength comes from MCTS amplifying the network, and the eval-sims sweep in the scientific log showed win-rate against the strong heuristic V4 rising with test-time simulations before plateauing. So V7 searches deeper during self-play (64 simulations — sharper policy/value targets) and much deeper at tournament time (128 simulations), with more gradient steps per game. This is the version that consistently beats the field — see the leaderboard.
lesson_7/policies/v7/__init__.py
"""V7 "AlphaStar" -- the headline policy: discounted-return self-play plus
SCALED SEARCH at both train and test time.
Conceptual breakthrough over V6: lean into search. AlphaGo Zero's strength comes
from MCTS amplifying the network; the eval-sims sweep (docs/scientific-log/
learning_0002.md) showed win-rate vs the strong heuristic V4 rising with more
test-time simulations before plateauing. V7 therefore searches deeper during
self-play (sharper policy/value targets) and much deeper at tournament time
(128 simulations), and takes more gradient steps per game. It keeps V6's
faster-win value discount. This is the version that consistently beats the field.
"""
from __future__ import annotations
from ..alphastar.train import Config
from ..alphastar.version import Version
VERSION = Version(
"V7 AlphaStar",
Config(
value_discount=0.99,
sims_selfplay=64,
sims_eval=128,
train_steps_per_game=8,
),
seed=0,
)
Nothing else changes: same network, same features, same MCTS, same training loop as
V5/V6 — only the Config numbers move. That is the point: in the AlphaGo-Zero
recipe, more search is a knob you can turn and reliably buy strength with
compute. The full code walk-through is at
How MicroStar works.