liuheyang0114 / olympicsproject Goto Github PK

Modified from https://github.com/jidiai/Competition_Olympics-Running.

• clean some code used for Jidi competitions
• new pipeline for developing

git clone https://github.com/Leo-xh/Competition_Olympics-Running.git
cd Competition_Olympics-Running

# training ppo with random opponent
python rl_trainer/main.py --device cuda --map 1  # 可以用不同的map，或者用--shuffle_map

# evaluating ppo with random opponent
python evaluation.py --my_ai ppo --my_ai_run_dir run1 --my_ai_run_episode 1500 --map 1  # run1可以改成自己的output directory

1. The random opponent may be too weak for developing new algorithms, you can implement other rule-based agents to compete with your algorithm.
2. You can also consider self-paly based training methods in training your agent.
3. For training a ppo algorithm, the given metrics may not be enough, you can add other metrics, e.g. clipping ratio, to help monitoring the training process.
4. Single-agent PPO may not work in difficult maps, and you should train your agent with --shuffle_map flag finally.