本文更新了南栖仙策为挑战赛提供的基线方案，对基线方案做了更详细说明，以及说明了如何评估和改善基线训练出的环境和模型策略。欢迎大家下载查阅。

1、基线方案基于Polixir Revive SDK 【 下载地址 】
2、基线方案详细文档见附件

3、基线代码，样例提交代码，以及用于指导参赛者学习一遍完整流程的jupyter notebook文件参见比赛所提供的starting_kit.zip

notebook的目录如下：

• Starting Kit Guidance
• Requirements
• Data organization
• Step 1: Derive user states from offline data
• Step 2: Learn a virtual environment
  
  • Evaluate the virtual environment
  • (a) Automatic evaluation with specific metrics
  • (b) Manual evaluation with histogram and rollout image
  • Get the model parameters for virtual environment
  • (a) Use the model with best metric
  • (b) Use the model with specific trail id
• Step 3: Learn a fair promotion policy from virtual environment
  
  • Environment and MDP Setup
  • (a) Action space
  • (b) Observation space
  • (c) Reward
  • Training of the policy
  • Evaluation of the policy
• Step 4: Generate a submission bundle
  
  • File structure
  • PolicyValidation file
  • Metadata
  • Create a submission