通过对手发现更好策略

本次运行结果

Dependency cache ready: /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/selfplay/.cache/online-experiments/hands-on-modern-rl-experiment04-board-selfplay-9becbb2cc824.ready
代码仓库： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/selfplay
实验运行目录： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/selfplay/modelscope-space/hands-on-modern-rl-experiment04-board-selfplay

可选任务：
  kuhn-poker           Kuhn Poker · kuhn_poker
  leduc-poker          Leduc Poker · leduc_poker
  tic-tac-toe          Tic-Tac-Toe · tic_tac_toe
  connect-four         Connect Four · connect_four

所选任务： Kuhn Poker
算法： CFR+
轮次安排：2 轮 × 250 步 = 500 总步数
每轮都会评估并保存一个可单独选择的模型。

Device: CPU
启动与在线 Studio 相同的训练流程…

Optional module pokerkit_wrapper was not importable: No module named 'pokerkit'

CFR+ iteration=250 exploitability=0.00041426
SAVE epoch=1/2 policy=policy.npz

CFR+ iteration=500 exploitability=0.00017270
SAVE epoch=2/2 policy=policy.npz

Saved 2 CFR policies and sampled one terminal game per epoch

最终阶段： complete
最终得分： -0.00017269684974360877
已保存 2 个本次运行可选择的轮次模型。
模型索引： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/selfplay/modelscope-space/hands-on-modern-rl-experiment04-board-selfplay/artifacts/notebook-models.json

运行下面的推理单元格，选择轮次模型并查看该策略的回放。

VBox(children=(Dropdown(description='训练轮次模型：', layout=Layout(width='95%'), options=(('轮次 2/2 · 500 步 · 得分 -0.0…

默认选择最新轮次；也可以选择其他已保存模型，然后点击 回放所选策略.

{
  "score": -0.00017269684974360877,
  "steps": 500,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "kuhn-poker-learned-policy.gif"
}

{
  "score": -0.00017269684974360877,
  "steps": 500,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "kuhn-poker-learned-policy.gif"
}