search

decisionforce / HACO

[ICLR 2022] Official implementation of paper: Efficient Learning of Safe Driving Policy via Human-AI Copilot Optimization
Apache License 2.0
42 stars 10 forks source link

Training considerations #8

Open xiaozhao12345 opened 5 months ago

xiaozhao12345 commented 5 months ago

Hello, are there any precautions when training haco, and how to control the frequency of manual intervention, because several problems were discovered during training:

  1. It is difficult to achieve the desired results during training
  2. Too much manual intervention will significantly reduce the performance of the system.
  3. After more than 300 rounds of training, the effect will also decrease. The main purpose is to understand when and how to intervene during training is better.