用 JAX 训练像素小游戏

本次运行结果

Dependency cache ready: /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/jax/.cache/online-experiments/hands-on-modern-rl-experiment07-jax-games-d3ef1adbd087.ready
代码仓库： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/jax
实验运行目录： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/jax/modelscope-space/hands-on-modern-rl-experiment07-jax-games

可选任务：
  asterix              Asterix MinAtar · Asterix-MinAtar
  breakout             Breakout MinAtar · Breakout-MinAtar
  freeway              Freeway MinAtar · Freeway-MinAtar
  space-invaders       Space Invaders MinAtar · SpaceInvaders-MinAtar

所选任务： Breakout MinAtar
算法： JAX REINFORCE
轮次安排：2 轮 × 50 步 = 100 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…

/Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/jax/_src/ops/scatter.py:105: FutureWarning: scatter inputs have incompatible types: cannot safely cast value from dtype=int32 to dtype=bool with jax_numpy_dtype_promotion=standard. In future JAX releases this will result in an error.
  warnings.warn(
/Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/jax/_src/ops/scatter.py:105: FutureWarning: scatter inputs have incompatible types: cannot safely cast value from dtype=float32 to dtype=bool with jax_numpy_dtype_promotion=standard. In future JAX releases this will result in an error.
  warnings.warn(

JIT compiling policy update · observation=(10, 10, 4) actions=3

JAX epoch=1/2 episode=50 loss=-0.010270 eval_return=1.600 replay_return=2.000
SAVE model=policy.npz replay=learned-policy.gif frames=18

JAX epoch=2/2 episode=100 loss=-0.010691 eval_return=0.000 replay_return=0.000
SAVE model=policy.npz replay=learned-policy.gif frames=6

Saved 2 independently selectable JAX policies and semantic replays

最终阶段： complete
最终得分： 0.0
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/jax/modelscope-space/hands-on-modern-rl-experiment07-jax-games/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 · 100 步 · 得分 0.00…

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": 0.0,
  "steps": 100,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "learned-policy.gif"
}

{
  "score": 0.0,
  "steps": 100,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "learned-policy.gif"
}