feat: 改进DQN训练稳定性和性能

Browse Source

- 将奖励裁剪替换为奖励缩放，保留奖励大小信号
- 添加学习率调度器，支持warmup和步进衰减
- 增加经验回放缓冲区大小至200,000
- 默认启用Dueling DQN架构
- 优化代码格式和参数传递
- 添加更多训练中间模型保存点

...

This commit is contained in:

刘航宇

2026-05-02 02:02:17 +08:00

parent 1c1cccd3f6

commit faf0d5ea42

12 changed files with 122 additions and 77 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

强化学习个人项目报告（Atari 游戏方向）/models/dqn_step_600000.pt

BIN

Binary file not shown.