agent.md

Build a brand new agent

1. Implement the step function in distar/agent/template/agent.py. This function should receive an observation and return a list of actions.

   the observation is a dict with 3 fields:

   • raw_obs: contains observation in form of protobuf (details).

   • opponent_obs: contains opponent's observation similar to raw_obs.

   • action_result: action result of last action the agent made(details).

   the action is a dict with 6 fields:

   • func_id: ability id from pysc2(details).

   • skip_steps: how many frames in game the agent doesn't need an observation, count from current frame.

   • queued: whether to put this action into queue(like human use shift).

   • unit_tags: a list contains unit tags which are selected units in this action.

   • target_unit_tag: a unit tag which is targetd unit in this action.

   • location: the target location in this action.

2. Change the agent's name from default to template at user_config.yaml

3. python -m distar.bin.play --game_type agent_vs_agent --model2 <agent you want to play against>

Modify our agent

1. Create a new branch named : git checkout -b <dev>
2. Modify any part of the default agent, here are some examples:

• Change the model size at distar/agent/default/model/actor_critic_default_config.yaml
• Change input features at distar/agent/default/lib/features.py
• Overwrite the agent's action in any certain situation at distar/agent/default/agent.py

3. Copy the directory of default agent from main branch and rename it main.
4. Change the agent's name from default to main at user_config.yaml
5. python -m distar.bin.play --game_type agent_vs_agent --model1 <agent you want to play against> --model2 <your own agent>