2 ~ Training and configuration

Training

In order to train the agent, execute the following command:

bash train.sh

This will automatically load a previous saved checkpoint, if one exists. Training progress is automatically saved at a configurable interval.

Configuration

The file A3C/options.py provides all the hyper-parameters values, which can be set to any value. Of particular note:

• state_generator: state generator to be used
• reward_generator: reward generator to be used
• parallel_size: number of threads
• steps_per_episode: maximum number of steps per episode
• match_count_for_evaluation: number of the most recent episodes to use for evaluating the agent
• save_interval_step: frequency of saves in training steps