数据之外的动作可靠吗

CPU · Python 3.12.13 · 4.21 s · 2026-10-22 08:53:43 UTC+8

本次运行结果

Behavior.true_return	-0.04669926315546036
Oracle.true_return	0.0

{'train_samples': 4800, 'test_samples': 1200, 'observed_action_range': [-0.2136465311050415, 0.18843042850494385]}

Trained three seeds without collecting additional environment rewards.

{
  "Naive Q": {
    "true_returns": [
      -0.13183709979057312,
      -0.21818819642066956,
      -0.057053759694099426
    ],
    "support_gaps": [
      0.5113945603370667,
      0.6254499554634094,
      0.3834943175315857
    ]
  },
  "Conservative": {
    "true_returns": [
      -0.05030469968914986,
      -0.04521385580301285,
      -0.04875820875167847
    ],
    "support_gaps": [
      0.021189402788877487,
      0.01570889726281166,
      0.013902311213314533
    ]
  },
  "Q + BC": {
    "true_returns": [
      -0.02849150076508522,
      -0.026607640087604523,
      -0.029490111395716667
    ],
    "support_gaps": [
      0.0489652082324028,
      0.05315358191728592,
      0.0440630204975605
    ]
  },
  "Behavior": {
    "true_return": -0.04669926315546036
  },
  "Oracle": {
    "true_return": 0.0
  }
}

Optional exercise: fill answer and enable RUN_EXERCISE.

metrics.json

{
  "Naive Q": {
    "true_returns": [
      -0.13183709979057312,
      -0.21818819642066956,
      -0.057053759694099426
    ],
    "support_gaps": [
      0.5113945603370667,
      0.6254499554634094,
      0.3834943175315857
    ]
  },
  "Conservative": {
    "true_returns": [
      -0.05030469968914986,
      -0.04521385580301285,
      -0.04875820875167847
    ],
    "support_gaps": [
      0.021189402788877487,
      0.01570889726281166,
      0.013902311213314533
    ]
  },
  "Q + BC": {
    "true_returns": [
      -0.02849150076508522,
      -0.026607640087604523,
      -0.029490111395716667
    ],
    "support_gaps": [
      0.0489652082324028,
      0.05315358191728592,
      0.0440630204975605
    ]
  },
  "Behavior": {
    "true_return": -0.04669926315546036
  },
  "Oracle": {
    "true_return": 0.0
  }
}