Files
Commercialization.tapadn/Tools/SmartLoadSensitivity/output/TapADN_智能预加载_敏感度验收报告.md

161 lines
7.8 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# TapADN 智能预加载策略验收报告(含随机偏好机器人 + 网络环境)
## 仿真前提
- 模型定位IAA 场景中 3 类广告位(激励/开屏/插屏)统一参与策略决策。
- 次留默认验收基线35%
- 随机机器人数量5000
- 场景进入->展示请求->预加载决策->展示时延为核心链路。
- 每次样本运行前先生成一批“偏好机器人”,再在其上分别运行:纯手动(无 smart与智能预加载两种模式。
- 预加载触发后在 cooldown 内有效一次,不命中将视作普通 load。
- `Immediate`:请求时已命中可直接播放的占比(值越高越好)
- `Waste`:预加载后在 cooldown 内未被消费即失效的比例(值越低越好)
- 网络环境Wi-Fi / 4G / 3G / 2G 按机器人偏好加权采样,逐回合独立变化。
## 机器人与网络设置
- 机器人类型随机采样以下偏好族reward_heavy、interstitial_focus、splash_driven、balanced、churn_sensitive、network_bound
- 每类机器人具有独立的场景进入偏好、请求偏好、fill 成功倍率、加载时延倍率和留存倍率。
- 网络环境以会话粒度采样低网速会同步影响网络请求率、fill 成功率和加载耗时。
## 参数扫描范围
- Threshold: `0.20~0.90` 步长 0.05
- Cooldown: `30,60,90,120,180,240,300`
- 次留:`20%,25%,30%,35%,40%,50%,60%`
## 基线对照
- 次留 35% 基线(纯手动,无 smart即时命中率0.00%
- 次留 35% 基线平均时延1615 msp95 2913
- 次留 35% 建议起点(当前模型):`threshold=0.25, cooldown=30s`
- 对比基线即时命中提升:`38.09%`
- 对比基线时延变化:`-565 ms`
- 平衡候选Waste<=12%
- threshold=0.25, cd=60sImmediate 25.12%Waste 4.88%
## 最优/最差(次留=35%
| 排名 | Threshold | Cooldown | 即时命中率 | 平均等待 | P95等待 | 浪费率 | 播放成功率 |
|---|---:|---:|---:|---:|---:|---:|---:|
| 1 | 0.25 | 30 | 38.09% | 1050 | 2787 | 14.03% | 81.35% |
| 2 | 0.20 | 30 | 36.65% | 1073 | 2790 | 13.96% | 81.35% |
| 3 | 0.25 | 60 | 25.12% | 1234 | 2774 | 4.88% | 77.96% |
| 4 | 0.20 | 60 | 24.81% | 1242 | 2830 | 6.75% | 77.45% |
| 5 | 0.30 | 30 | 24.65% | 1285 | 2894 | 7.73% | 77.79% |
| 6 | 0.30 | 60 | 24.56% | 1239 | 2812 | 5.76% | 77.61% |
| 7 | 0.35 | 30 | 23.05% | 1318 | 2895 | 5.44% | 77.58% |
| 8 | 0.50 | 30 | 22.96% | 1320 | 2893 | 4.99% | 77.54% |
### 最差 8 组
| 排名 | Threshold | Cooldown | 即时命中率 | 平均等待 | P95等待 | 浪费率 | 播放成功率 |
|---|---:|---:|---:|---:|---:|---:|---:|
| 1 | 0.90 | 300 | 0.00% | 1629 | 2924 | 0.00% | 70.55% |
| 2 | 0.90 | 240 | 0.00% | 1630 | 2899 | 0.00% | 70.09% |
| 3 | 0.90 | 180 | 0.00% | 1622 | 2907 | 0.00% | 70.60% |
| 4 | 0.90 | 120 | 0.00% | 1633 | 2943 | 0.00% | 70.51% |
| 5 | 0.90 | 90 | 0.00% | 1637 | 2963 | 0.00% | 70.84% |
| 6 | 0.90 | 60 | 0.00% | 1628 | 2863 | 0.00% | 70.72% |
| 7 | 0.90 | 30 | 0.00% | 1627 | 2933 | 0.00% | 69.52% |
| 8 | 0.85 | 300 | 0.00% | 1623 | 2910 | 0.00% | 70.69% |
## 与纯手动模式对比次留35%
- 纯手动Immediate `0.00%`AvgWait `1615`msWaste `0.00%`
- smart 最优:阈值 `0.25`cd `30`Immediate `38.09%`AvgWait `1050`msWaste `14.03%`
- 增益:`38.09%`
### 网络分层对比次留35%
- WIFI手动即时 `0.00%` / 等待 `1248ms`;智能(`threshold=0.25, cd=30s`)即时 `40.13%` / 等待 `767ms`;提升 `40.13%`
- 4G手动即时 `0.00%` / 等待 `1601ms`;智能(`threshold=0.25, cd=30s`)即时 `39.66%` / 等待 `1003ms`;提升 `39.66%`
- 3G手动即时 `0.00%` / 等待 `2165ms`;智能(`threshold=0.20, cd=30s`)即时 `32.64%` / 等待 `1498ms`;提升 `32.64%`
- 2G手动即时 `0.00%` / 等待 `3044ms`;智能(`threshold=0.25, cd=30s`)即时 `32.06%` / 等待 `2116ms`;提升 `32.06%`
## 次留敏感度35%基线)
- 次留 20%:手动基线时延 1618ms智能最佳 `threshold=0.20, cd=30s`,最佳 Immediate 38.69%Waste 13.14%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 25%:手动基线时延 1641ms智能最佳 `threshold=0.20, cd=30s`,最佳 Immediate 37.52%Waste 14.16%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 30%:手动基线时延 1629ms智能最佳 `threshold=0.20, cd=30s`,最佳 Immediate 37.42%Waste 13.33%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 35%:手动基线时延 1615ms智能最佳 `threshold=0.25, cd=30s`,最佳 Immediate 38.09%Waste 14.03%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 40%:手动基线时延 1617ms智能最佳 `threshold=0.20, cd=30s`,最佳 Immediate 38.05%Waste 13.64%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 50%:手动基线时延 1618ms智能最佳 `threshold=0.25, cd=30s`,最佳 Immediate 38.02%Waste 14.06%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
- 次留 60%:手动基线时延 1619ms智能最佳 `threshold=0.25, cd=30s`,最佳 Immediate 38.01%Waste 14.88%;最差 `threshold=0.90, cd=300s`Immediate 0.00%Waste 0.00%。
## 交叉参数观察
- 在同一保留率下,阈值下调能显著提高 `Immediate`,但通常也抬高 `Waste`
- cooldown 拉长可降低 waste减少重复预加载/空耗),但可能提高用户等待。
- 在弱网(尤其低分配机器人更多时)场景,建议提高 waste 上限约束后再考虑 lower threshold。
## 图表示例(输出文件)
- `heatmap_immediate_r_0.20.png`
![](heatmap_immediate_r_0.20.png)
- `heatmap_immediate_r_0.35.png`
![](heatmap_immediate_r_0.35.png)
- `heatmap_immediate_r_0.60.png`
![](heatmap_immediate_r_0.60.png)
- `heatmap_preload_success_r_0.20.png`
![](heatmap_preload_success_r_0.20.png)
- `heatmap_preload_success_r_0.35.png`
![](heatmap_preload_success_r_0.35.png)
- `heatmap_preload_success_r_0.60.png`
![](heatmap_preload_success_r_0.60.png)
- `heatmap_wait_ms_r_0.20.png`
![](heatmap_wait_ms_r_0.20.png)
- `heatmap_wait_ms_r_0.35.png`
![](heatmap_wait_ms_r_0.35.png)
- `heatmap_wait_ms_r_0.60.png`
![](heatmap_wait_ms_r_0.60.png)
- `heatmap_waste_ratio_r_0.20.png`
![](heatmap_waste_ratio_r_0.20.png)
- `heatmap_waste_ratio_r_0.35.png`
![](heatmap_waste_ratio_r_0.35.png)
- `heatmap_waste_ratio_r_0.60.png`
![](heatmap_waste_ratio_r_0.60.png)
- `line_immediate_vs_retention.png`
![](line_immediate_vs_retention.png)
- `line_mode_immediate_vs_retention.png`
![](line_mode_immediate_vs_retention.png)
- `line_mode_wait_vs_retention.png`
![](line_mode_wait_vs_retention.png)
- `line_network_immediate_vs_retention.png`
![](line_network_immediate_vs_retention.png)
- `line_network_wait_vs_retention.png`
![](line_network_wait_vs_retention.png)
- `line_threshold_0.55_retention_0.20.png`
![](line_threshold_0.55_retention_0.20.png)
- `line_threshold_0.55_retention_0.25.png`
![](line_threshold_0.55_retention_0.25.png)
- `line_threshold_0.55_retention_0.30.png`
![](line_threshold_0.55_retention_0.30.png)
- `line_threshold_0.55_retention_0.35.png`
![](line_threshold_0.55_retention_0.35.png)
- `line_threshold_0.55_retention_0.40.png`
![](line_threshold_0.55_retention_0.40.png)
- `line_threshold_0.55_retention_0.50.png`
![](line_threshold_0.55_retention_0.50.png)
- `line_threshold_0.55_retention_0.60.png`
![](line_threshold_0.55_retention_0.60.png)
- `line_wait_vs_retention.png`
![](line_wait_vs_retention.png)
## 验收结论
- 建议首轮灰度点:`threshold=0.25`, `cooldown=60s`(兼顾即时性与 waste
- 次留 35% 下智能策略最佳立即命中来自 `threshold=0.25, cd=30s`,即时率 `38.09%`Waste `14.03%`
- 建议将 `smart_preload` 配置作为实验变量:先以纯手动为对照,再按次留分层和网络监控逐步放量。